On Mon, Feb 23, 2009 at 01:01:05PM +0530, arun c wrote: > Hi, ... > > I suspect you need/want at least one wmb() between those three steps > > depending on what the remote host is polling. > > The "VALID" field sounds redundant to me and I would drop it. > > Updating the write_index should be enough clue for the remote side > > to know which indexs are valid. > > This is useful for me when read_index = write_index > It will stop host to overwrite the buffer when there > is unprocessed commands at the target side. Keeping track of indexes is more efficient if the SW can process multiple requests in a single pass....So I prefer it over using a "VALID" bit. > Target never reads write_index here. That's what I had guessed earlier. If the Target polled the write_index instead of VALID bit, the VALID bit wouldn't be needed. It might be nice to have for debugging but could be ignored later. > > >> Target reads commands at read_index, > >> if and only if valid field is VALID and > >> after processing it makes valid field > >> INVALID and increments read_index. > > > > So "VALID" really means "BUSY" or "INFLIGHT". > > > > Don't really need share both VALID bit and head/tail indexes. > > Avoiding sharing of the index could conserve a fair amount of CPU cycles > > on the host side. > > I really don't understand this, for me host and target > is sharing a buffer sitting at the PCI space. I saw your follow up. The buffer needs to be shared. BTW, my point is the MMIO reads are expensive for the host. Use "get_cycles()" on the host side to see how many cycles each MMIO read takes to access the SDRAM on the target (across the PCI bus). hth, grant > The target is exposing its SDRAM over PCI and > the buffers are local to target and over PCI to host. > > In this case how to avoid sharing? > > > > >> It seems to be working for me > >> I need to investigate further for any > >> cache coherency issues, write now > >> I am running with dcache off. > > > > This is arch specific. Host side is usually cache coherent. > > IIRC, some embedded PPC are not. > > > Regards, > Arun C -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
