Stephan Szabo <sszabo@megazone.bigpanda.com> writes:
> On Tue, 24 Aug 2004, Christopher Kings-Lynne wrote:
>> OK, what I mean is to know if a row is locked by any backend, why can't
>> we just put a reference count of the number of locks on that row,
>> instead of recording each backend separately? Wouldn't that require a
>> fixed amount of shared mem?
> AFAICT you have to do something on top of that to allow deadlock
> detection. If transaction X has a shared row lock on A and is waiting on
> a lock for me and I want to get an exclusive row lock on A, how do I
> detect that it's a deadlock?
I think the speed complaint I was just raising could possibly be
answered by setting an infomask bit indicating that the row might
be present in a separate table of active row locks. (I'm not sure
how the bit would get cleared without race conditions, but let's
suppose that can be done.) A little hashing, a little spill-to-disk
logic, and it might be done. But that's just handwaving... anyone
want to try to fill in the details?
[ But to answer Chris' question: no, I don't see any way that this
could be a fixed-size table. You will need that spill-to-disk bit. ]
regards, tom lane