On 10-10-23, 21:23, Kelvin.Cao@xxxxxxxxxxxxx wrote: > On Mon, 2023-10-09 at 11:08 +0530, Vinod Koul wrote: > > > u64 size_to_transfer; > > > > Why cant the client driver write to doorbell, is there anything which > > prevents us from doing so? > > I think the potential challenge here for the client driver to ring db > is that the client driver (host RC) is a different requester in the > PCIe hierarchy compared to DMA EP, in which case PCIe ordering need to > be considered. > > As PCIe ensures that reads don't pass writes, we can insert a read DMA > operation with DMA_PREP_FENSE flag in between the two DMA writes (one > for data transfer and one for notification) to ensure the ordering for > the same requester DMA EP. I'm not sure if the RC could ensure the same > ordering if the client driver issue MMIO write to db after the data DMA > and read DMA completion, so that the consumer is guaranteed the > transferred data is ready in memory when the db is triggered by the > client MMIO write. I guess it's still doable with MMIO write but just > some special consideration needed. Given that it is a single value, overhead of doing a new txn would be higher than a mmio write! I think that should be preferred -- ~Vinod
