BUG #19836: Backend crash with large hash partition modulus
BUG #19836: Backend crash with large hash partition modulus
От:
PG Bug reporting form <noreply@postgresql.org>
Дата:
The following bug has been logged on the website: Bug reference: 19836 Logged by: Sumedh Pathak Email address: sumedhpathak@gmail.com PostgreSQL version: 19rc1 Operating system: MacOS Description: Hi, The following SQL crashes PostgreSQL 20devel (72b9498df9a, macOS ARM64): CREATE TABLE parent (i integer) PARTITION BY HASH (i); CREATE TABLE p0 PARTITION OF parent FOR VALUES WITH (MODULUS 3, REMAINDER 0); CREATE TABLE p1 PARTITION OF parent FOR VALUES WITH (MODULUS 2147483646, REMAINDER 2); The third statement causes a crash and server recovery. In check_new_partition_bound(), remainder += spec->modulus overflows signed int (2 + 2147483646), leading to an out-of-bounds lookup. The crash was verified on PG 17,18 as well, but it seems like it exists since PG11. Thanks.
Re: BUG #19836: Backend crash with large hash partition modulus
От:
David Rowley <dgrowleyml@gmail.com>
Дата:
On Sun, 8 Nov 2026 at 04:54, PG Bug reporting form
wrote:
> The following SQL crashes PostgreSQL 20devel (72b9498df9a, macOS ARM64):
> CREATE TABLE parent (i integer) PARTITION BY HASH (i);
> CREATE TABLE p0 PARTITION OF parent
> FOR VALUES WITH (MODULUS 3, REMAINDER 0);
> CREATE TABLE p1 PARTITION OF parent
> FOR VALUES WITH (MODULUS 2147483646, REMAINDER 2);
Thanks for the report.
It seems fine to just do:
@@ -2919,8 +2919,8 @@ check_new_partition_bound(char *relname, Relation parent,
if (partdesc->nparts > 0)
{
- int
greatest_modulus;
- int remainder;
+ int64 greatest_modulus;
+ int64 remainder;
That'll prevent the wrap and allow the do/while loop to terminate when
remainder is >= greatest_modulus. Bumping greatest_modulus isn't
really required, but I did that because I thought it was neater to
keep them the same size.
I can patch this when the freeze is over.
David
Re: BUG #19836: Backend crash with large hash partition modulus
От:
David Rowley <dgrowleyml@gmail.com>
Дата:
On Mon, 9 Nov 2026 at 13:10, David Rowley wrote:
> @@ -2919,8 +2919,8 @@ check_new_partition_bound(char *relname, Relation parent,
>
> if (partdesc->nparts > 0)
> {
> - int
> greatest_modulus;
> - int remainder;
> + int64 greatest_modulus;
> + int64 remainder;
>
> That'll prevent the wrap and allow the do/while loop to terminate when
> remainder is >= greatest_modulus. Bumping greatest_modulus isn't
> really required, but I did that because I thought it was neater to
> keep them the same size.
>
> I can patch this when the freeze is over.
Pushed.
David