table_open/table_close with different lock mode

On Fri, Jul 21, 2023 at 02:05:56PM +0800, Junwang Zhao wrote:

I noticed there are some places calling table_open with
RowExclusiveLock but table_close with NoLock, like in function
toast_save_datum.

Can anybody explain the underlying logic, thanks in advance.

This rings a bell. This is a wanted behavior, see commit f99870d and
its related thread:
/messages/by-id/17268-d2fb426e0895abd4@postgresql.org

The tests added by this commit in src/test/isolation/ will show the
difference in terms of the way the toast values get handled with and
without the change.
--
Michael

On Fri, Jul 21, 2023 at 2:26 PM Michael Paquier <michael@paquier.xyz> wrote:

On Fri, Jul 21, 2023 at 02:05:56PM +0800, Junwang Zhao wrote:

I noticed there are some places calling table_open with
RowExclusiveLock but table_close with NoLock, like in function
toast_save_datum.

Can anybody explain the underlying logic, thanks in advance.

This rings a bell. This is a wanted behavior, see commit f99870d and
its related thread:
/messages/by-id/17268-d2fb426e0895abd4@postgresql.org

I see this patch, so all the locks held by a transaction will be released
at the commit phase, right? Can you show me where the logic is located?

The tests added by this commit in src/test/isolation/ will show the
difference in terms of the way the toast values get handled with and
without the change.
--
Michael

--
Regards
Junwang Zhao

On Thu, Jul 20, 2023 at 11:38 PM Junwang Zhao <zhjwpku@gmail.com> wrote:

On Fri, Jul 21, 2023 at 2:26 PM Michael Paquier <michael@paquier.xyz> wrote:

On Fri, Jul 21, 2023 at 02:05:56PM +0800, Junwang Zhao wrote:

I noticed there are some places calling table_open with
RowExclusiveLock but table_close with NoLock, like in function
toast_save_datum.

Can anybody explain the underlying logic, thanks in advance.

This rings a bell. This is a wanted behavior, see commit f99870d and
its related thread:
/messages/by-id/17268-d2fb426e0895abd4@postgresql.org

I see this patch, so all the locks held by a transaction will be released
at the commit phase, right? Can you show me where the logic is located?

The NoLock is simple a marker that tells the underlying machinery to
not bother releasing any locks. As a matter of fact, you can pass
NoLock in *_open() calls, too, to indicate that you don't want any new
locks, perhaps because the transaction has already taken an
appropriate lock on the object.

As for lock-releasing codepath at transaction end, see
CommitTransaction() in xact.c, and specifically at the
ResourceOwnerRelease() calls in there.

Best regards,
Gurjeet
http://Gurje.et

On Fri, Jul 21, 2023 at 2:57 PM Gurjeet Singh <gurjeet@singh.im> wrote:

On Thu, Jul 20, 2023 at 11:38 PM Junwang Zhao <zhjwpku@gmail.com> wrote:

On Fri, Jul 21, 2023 at 2:26 PM Michael Paquier <michael@paquier.xyz> wrote:

On Fri, Jul 21, 2023 at 02:05:56PM +0800, Junwang Zhao wrote:

I noticed there are some places calling table_open with
RowExclusiveLock but table_close with NoLock, like in function
toast_save_datum.

Can anybody explain the underlying logic, thanks in advance.

This rings a bell. This is a wanted behavior, see commit f99870d and
its related thread:
/messages/by-id/17268-d2fb426e0895abd4@postgresql.org

I see this patch, so all the locks held by a transaction will be released
at the commit phase, right? Can you show me where the logic is located?

The NoLock is simple a marker that tells the underlying machinery to
not bother releasing any locks. As a matter of fact, you can pass
NoLock in *_open() calls, too, to indicate that you don't want any new
locks, perhaps because the transaction has already taken an
appropriate lock on the object.

As for lock-releasing codepath at transaction end, see
CommitTransaction() in xact.c, and specifically at the
ResourceOwnerRelease() calls in there.

Great, thanks for the thorough explanation, will look into the code :)

Best regards,
Gurjeet
http://Gurje.et

--
Regards
Junwang Zhao