Tom Lane wrote:
>
> 2. I made most opens of system relations grab AccessShareLock if
> read-only, or RowExclusiveLock if read-write, on the theory that ^^^^^^^^^^^^^^^^
> these accesses correspond to an ordinary search or update of a user
> relation. This maximizes concurrency of access to the system tables.
There are problems here. In the case of normal UPDATE/DELETE
(when RowExclusiveLock is acquired) Executor takes care about
the-same-row writers, but other parts of system don't check
is tuple read being updated concurrent transaction or not.
This is the old bug (pre-6.5.X released WRITE lock just after
system table was modified). I had no time to fix it and so
just changed old WRITE lock with new AccessExclusiveLock.
But we have to handle this in proper way (wait if t_xmax
is id of an active transaction).
Vadim