On Tue, Aug 10, 2021 at 4:15 PM Peter Feiner <pfeiner@xxxxxxxxxx> wrote: > > Races to increment by a few pages might be OK, so long as imprecision > > is acceptable, but decrementing by an entire bitmap could cause the > > stat to get waaay off from the state of the VM. > > My original use case was to know the rate at which memory was being > dirtied to predict how expensive a live migration would be. I didn't > need full precision, but I would have needed a bound on the slop. A > "few pages" isn't a bound :-) I think the agreement with VM-scoped statistics is that slop is OK, better than paying the cost of locking/atomics. If we want to be exact, it'd have to be a vCPU stat. > IMO, this patch isn't worth figuring out without a use case. It's > complex and has perf overhead. Maybe just drop it? Knowing the approximate rate at which pages are being dirtied would be a nice-to-have for debugging, IMO. Just treating this stat as monotonic would greatly simplify it and avoid the overhead. -- Thanks, Oliver
