Horton
Menu

Thread

  1. It sorta works, but I'm confused about locking

    Tom Lane <tgl@sss.pgh.pa.us> — 1998-09-27T21:34:33Z

    I've got this new notify code almost working, but...

What exactly is the protocol for locking a table that you intend to
modify?  The old notify code just did RelationSetLockForWrite(),
but it's not clear to me that that works correctly --- for one thing,
it doesn't seem to offer any way of detecting failure to acquire the
lock.  (RelationSetLockForWrite calls MultiLockReln, which *does*
return a boolean, but RelationSetLockForWrite ignores it...)  Also,
it's not at all clear whether I should call RelationUnsetLockForWrite
at the end of the routine or not; some existing code does, some doesn't.

I'm concerned because interlocking of the specialized NOTIFY-related
statements seems to work fine, but they seem not to be interlocked
against user operations on the pg_listener table.

For example, this works as I'd expect:

	psql#1				psql#2

	begin;
	listen z;

					notify z;
					(hangs up until #1 commits)

	end;

because "listen" acquires a write lock on the pg_listener table, which
the notify has to wait for.

But this doesn't work as I'd expect:

	psql#1				psql#2

	begin;
	select * from pg_listener;

					notify z;
					(completes immediately)

	end;

Seems to me the "select" should acquire a read lock that would prevent
the #2 backend from writing pg_listener until the end of #1's transaction.

Is there a bug here, or is there some special definition of user access
to a system table that means the select isn't acquiring a read lock?
Selects and updates on ordinary user tables seem to interlock fine...

			regards, tom lane

  2. Re: [HACKERS] It sorta works, but I'm confused about locking

    Massimo Dal Zotto <dz@cs.unitn.it> — 1998-09-28T16:46:58Z

    > 
> I've got this new notify code almost working, but...

Could you plase send it to me ?

> What exactly is the protocol for locking a table that you intend to
> modify?  The old notify code just did RelationSetLockForWrite(),
> but it's not clear to me that that works correctly --- for one thing,
> it doesn't seem to offer any way of detecting failure to acquire the
> lock.  (RelationSetLockForWrite calls MultiLockReln, which *does*
> return a boolean, but RelationSetLockForWrite ignores it...)  Also,
> it's not at all clear whether I should call RelationUnsetLockForWrite
> at the end of the routine or not; some existing code does, some doesn't.

It is not done where there is an immediate CommitTransactionCommand which
already releases the locks.

> I'm concerned because interlocking of the specialized NOTIFY-related
> statements seems to work fine, but they seem not to be interlocked
> against user operations on the pg_listener table.
> 
> For example, this works as I'd expect:
> 
> 	psql#1				psql#2
> 
> 	begin;
> 	listen z;
> 
> 					notify z;
> 					(hangs up until #1 commits)
> 
> 	end;
> 
> because "listen" acquires a write lock on the pg_listener table, which
> the notify has to wait for.
> 
> But this doesn't work as I'd expect:
> 
> 	psql#1				psql#2
> 
> 	begin;
> 	select * from pg_listener;
> 
> 					notify z;
> 					(completes immediately)
> 
> 	end;
> 
> Seems to me the "select" should acquire a read lock that would prevent
> the #2 backend from writing pg_listener until the end of #1's transaction.
> 
> Is there a bug here, or is there some special definition of user access
> to a system table that means the select isn't acquiring a read lock?
> Selects and updates on ordinary user tables seem to interlock fine...

Very strange, it seems a bug. But users should not know about pg_listener.

-- 
Massimo Dal Zotto

+----------------------------------------------------------------------+
|  Massimo Dal Zotto                email:  dz@cs.unitn.it             |
|  Via Marconi, 141                 phone:  ++39-461-534251            |
|  38057 Pergine Valsugana (TN)     www:  http://www.cs.unitn.it/~dz/  |
|  Italy                            pgp:  finger dz@tango.cs.unitn.it  |
+----------------------------------------------------------------------+

  3. Re: [HACKERS] It sorta works, but I'm confused about locking

    Bruce Momjian <maillist@candle.pha.pa.us> — 1998-10-01T21:29:42Z

    > I've got this new notify code almost working, but...
> 
> What exactly is the protocol for locking a table that you intend to
> modify?  The old notify code just did RelationSetLockForWrite(),
> but it's not clear to me that that works correctly --- for one thing,
> it doesn't seem to offer any way of detecting failure to acquire the
> lock.  (RelationSetLockForWrite calls MultiLockReln, which *does*
> return a boolean, but RelationSetLockForWrite ignores it...)  Also,
> it's not at all clear whether I should call RelationUnsetLockForWrite
> at the end of the routine or not; some existing code does, some doesn't.

Welcome to PostgreSQL.  This is the type of thing I saw when doing the
mega-patch, were people were not doing things consistently, because some
coders do not understand what is happening inside the code, and just
write something that works, sort of.

I recommend you check out what is currently done properly, and fix the
ones that are incorrect.

I can imagine some cases where you would want to get a lock and keep it
until the end of the transaction, and other times when you would want to
release it before transaction end.


> 
> I'm concerned because interlocking of the specialized NOTIFY-related
> statements seems to work fine, but they seem not to be interlocked
> against user operations on the pg_listener table.
> 
> For example, this works as I'd expect:
> 
> 	psql#1				psql#2
> 
> 	begin;
> 	listen z;
> 
> 					notify z;
> 					(hangs up until #1 commits)
> 
> 	end;
> 
> because "listen" acquires a write lock on the pg_listener table, which
> the notify has to wait for.
> 
> But this doesn't work as I'd expect:
> 
> 	psql#1				psql#2
> 
> 	begin;
> 	select * from pg_listener;
> 
> 					notify z;
> 					(completes immediately)
> 
> 	end;
> 
> Seems to me the "select" should acquire a read lock that would prevent
> the #2 backend from writing pg_listener until the end of #1's transaction.
> 
> Is there a bug here, or is there some special definition of user access
> to a system table that means the select isn't acquiring a read lock?
> Selects and updates on ordinary user tables seem to interlock fine...

Select certainly should be locking.  Something is wrong, but I am not
sure what.  If you want me to check into it, let me know.


-- 
  Bruce Momjian                        |  http://www.op.net/~candle
  maillist@candle.pha.pa.us            |  (610) 853-3000
  +  If your life is a hard drive,     |  830 Blythe Avenue
  +  Christ can be your backup.        |  Drexel Hill, Pennsylvania 19026