Re: [HACKERS] Lock freeze ? in MVCC

OK, let me comment on this. It does not to see this as a deadlock
because session 3 really doesn't have a lock at the point it is hanging.
A deadlock would be if 1 has a lock that 3 is waiting for, and 3 has a
lock 1 is waiting for.

Hold on, I think I see what you are saying now. It seems the locking
code assume table-level locking, while the new code now has MVCC. I
better look at this. This could be ugly to fix. I look for matching
lock structure pointers in different backends(lock.c), but now I see
that 1 and 2 both are waiting for table tt, but they have different
locks structures, because they are different types of locks. Yikes.
Maybe I can hack something in there, but I can't imagine how yet. Maybe
Vadim will have a hint.

---------------------------------------------------------------------------

session-1 => create table tt (id int4);

session-1 => begin;
session-1 => insert into tt values (1);

session-2 => begin;
session-2 => insert into tt values (2);

session-3 => begin;
session-3 => lock table tt;
(blocked)

session-1 => update tt set id=1 where id=1;
(blocked)

session-2 => end;

session-2 returns immediately,but session-3 and session-1
are still blocked

This phenomenon seems to be caused by LockResolveCon
flicts() or DeadLockCheck().
Both session-1 and session-2 acquire RowExclusive locks
by insert operations(InitPlan() in execMain.c).
The AccessExclusive lock of session-3 is queued waiting
for the release of above locks.
When the update operation of session-1 is executed,the
second RowExclusive lock is rejected by LockResolve
Conflicts() and queued after the AccessExclusive lock
of session-3.
The state is like deadlock but DeadLockCheck() doesn't
regard the state as deadlock.

Thanks.

Hiroshi Inoue
Inoue(at)tpf(dot)co(dot)jp

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