On Nov 3 10:18, Jinhao Fan wrote: > at 7:10 PM, Klaus Jensen <i...@irrelevant.dk> wrote: > > > This doesn't do what you expect it to. By not updaring the eventidx it > > will fall behind the actual head, causing the host to think that the > > device is not processing events (but it is!), resulting in doorbell > > ringing. > > I’m not sure I understand this correctly. > > In 7.13.1 in NVMe Spec 1.4c it says "If updating an entry in the Shadow > Doorbell buffer **changes** the value from being less than or equal to the > value of the corresponding EventIdx buffer entry to being greater than that > value, then the host shall also update the controller's corresponding > doorbell register to match the value of that entry in the Shadow Doorbell > buffer.” > > So my understanding is that once the eventidx falls behind the actual head, > the host will only ring the doorbell once but *not* for future submissions. > > Is this not what real hosts are doing?
I agree that the spec is a little unclear on this point. In any case, in Linux, when the driver has decided that the sq tail must be updated, it will use this check: (new_idx - event_idx - 1) < (new_idx - old) So it doesn't account for if or not eventidx was already behind.
signature.asc
Description: PGP signature
