[PATCH] LockAcquireExtended improvement

Hi hackers,

I found a problem when doing the test shown below:

Time

Session A

Session B


T1

postgres=# create table test(a int);

CREATE TABLE

postgres=# insert into test values (1);

INSERT 0 1

 


T2

postgres=# begin;                                     

BEGIN                                                 

postgres=*# lock table test in access exclusive mode ;

LOCK TABLE                                            

 


T3

 

postgres=# begin;

BEGIN

postgres=*# lock table test in exclusive mode ;


T4

Case 1:

postgres=*# lock table test in share row exclusive mode nowait;

ERROR:  could not obtain lock on relation "test"               

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

Case 2:

postgres=*# lock table test in share row exclusive mode;

LOCK TABLE

 


At T4 moment in session A, (case 1) when executing SQL “lock table test in share row exclusive mode nowait;”, an error occurs with message “could not obtain lock on relation test";However, (case 2) when executing the SQL above without nowait, lock can be obtained successfully.

Digging into the source code, I find that in case 2 the lock was obtained in the function ProcSleep instead of LockAcquireExtended. Due to nowait logic processed before WaitOnLock->ProcSleep, acquiring lock failed in case 1. Can any changes be made so that the act of such lock granted occurs before WaitOnLock?

 

Providing a more universal case:

Transaction A already holds an n-mode lock on table test. If then transaction A requests an m-mode lock on table Test, m and n have the following constraints:

(lockMethodTable->conflictTab[n] & lockMethodTable->conflictTab[m]) == lockMethodTable->conflictTab[m]

Obviously, in this case, m<=n.

Should the m-mode lock be granted before WaitOnLock?

&nbsp;

In the case of m=n (i.e. we already hold the lock), the m-mode lock is immediately granted in the LocalLock path, without the need of lock conflict check.

Based on the facts above, can we obtain a weaker lock (m<n) on the same object within the same transaction without doing lock conflict check?

Since m=n works, m<n should certainly work too.

&nbsp;

I am attaching a patch here with which the problem in case 1 fixed.

With&nbsp;Regards,
Jingxian Li.

&nbsp;