Tom Lane escribió:
> Alvaro Herrera <alvherre@commandprompt.com> writes:
> > Tom Lane escribió:
> >> I think the fundamental bug here is that you tried to skip using the
> >> ResourceOwner mechanism for snapshot references. That's basically
> >> not gonna work.
>
> > Right :-( I'll see how to go about this.
>
> It strikes me that you might be able to remove the registered-snapshot
> infrastructure in snapmgr.c, and end up with about the same amount of
> code overall, just with the responsibility in resowner.c.
Seems to work, and fixes Pavan test case as well.
$ runpg 00head showdiff | diffstatbackend/utils/resowner/resowner.c | 99
++++++++++++++++++++backend/utils/time/snapmgr.c | 180
++++++-------------!!!!!!!!!!!!!!!!!include/utils/resowner.h | 8 +include/utils/snapmgr.h | 3 4
fileschanged, 143 insertions(+), 59 deletions(-), 88 modifications(!)
I need to fix some comments before publishing the patch.
The only thing I'm now missing is SnapshotResetXmin(). It currently
looks like this:
static void
SnapshotResetXmin(void)
{if (RegisteredSnapshotList == NULL && ActiveSnapshot == NULL) MyProc->xmin = InvalidTransactionId;
}
After the patch we don't have any way to detect whether resowner.c has
any snapshot still linked to. I assume there's no objection to adding a
new entry point in resowner.c for this.
--
Alvaro Herrera http://www.CommandPrompt.com/
The PostgreSQL Company - Command Prompt, Inc.