Error exits (Re: [HACKERS] Open 6.5 items)

Vadim Mikheev <vadim@krs.ru> writes:
>> It seems elog(FATAL) doesn't release allocated buffer pages.
>> It's OK ?
>> AFAIC elog(FATAL) causes proc_exit(0) and proc_exit() doesn't
>> call ResetBufferPool().

> Seems to me that elog(FATAL) should call siglongjmp(Warn_restart, 1),
> like elog(ERROR), but force exit in tcop main loop after
> AbortCurrentTransaction(). AbortTransaction() does pretty nice
> things like RelationPurgeLocalRelation(false) and DestroyNoNameRels()...

Seems reasonable to me.  It seems to me that elog(FATAL) means "this
backend is too messed up to continue, but I think the rest of the
backends can keep going."  So we need to clean up our allocated
resources before quitting.  abort() is for the cases where we think
shared memory may be corrupted and everyone must bail out.  We might
need to revisit the uses of each routine and make sure that different
error conditions are properly classified.

Of course, if things *are* messed up then trying to AbortTransaction
might make it worse.  How bad is that risk?
        regards, tom lane