On 2021-01-19 14:16:07 +0800, Craig Ringer wrote: > AFAICS it'd be necessary to expand PROCLOG to expose this in shmem. > Probably by adding a small bitfield where bit 0 is set if there's a txn > level lock and bit 1 is set if there's a session level lock. But I'm not > convinced that expanding PROCLOCK is justifiable for this. sizeof(PROCLOCK) > is 64 on a typical x64 machine. Adding anything to it increases it to 72 > bytes.
Indeed - I really don't want to increase the size, it's already a problem.
> It's frustrating to be unable to tell the difference between session-level > and txn-level locks in diagnostic output.
It'd be useful, I agree.
> And the deadlock detector has no way to tell the difference when > selecting a victim for a deadlock abort - it'd probably make sense to > prefer to send a deadlock abort for txn-only lockers.
I'm doubtful this is worth going for.
> But I'm not sure I see a sensible way to add the info - PROCLOCK is > already free of any padding, and I wouldn't want to use hacks like > pointer-tagging.
I think there's an easy way to squeeze out space: make groupLeader be an integer index into allProcs instead. That requires only 4 bytes...
Alternatively, I think it'd be reasonably easy to add the scope as a bit in LOCKMASK - there's plenty space.
I was wondering about that, but concerned that there would be impacts I did not understand.
I'm happy to pursue that angle.
Just so this thread isn't left dangling, I'm just not going to get time to follow up on this work with a concrete patch and test suite change.
If anyone else later on wants to differentiate between session and txn LWLocks they could start with the approach proposed here.