On Mon, Jan 8, 2018 at 1:53 PM, Javier González <jg@xxxxxxxxxxx> wrote: >> On 8 Jan 2018, at 12.54, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: >> >> On Fri, Jan 05, 2018 at 07:33:36PM +0100, Matias Bjørling wrote: >>> On 01/05/2018 04:42 PM, Jens Axboe wrote: >>>> On Fri, Jan 05 2018, Matias Bjørling wrote: >>>>> From: Javier González <javier@xxxxxxxxxxxx> >>>>> >>>>> Since pblk registers its own block device, the iostat accounting is >>>>> not automatically done for us. Therefore, add the necessary >>>>> accounting logic to satisfy the iostat interface. >>>> >>>> Ignorant question - why is it a raw block device, not using blk-mq? >>> >>> The current flow is using the raw block device, together with the blk-mq >>> nvme device driver. A bio is sent down to the nvme_nvm_submit_io() path in >>> the /drivers/nvme/host/lightnvm.c file. From there it attaches the to NVMe >>> blk-mq implementation. >>> >>> Is there a better way to do it? >> >> I suspect the right way to do things is to split NVMe for different >> I/O command sets, and make this an I/O command set. > > This makes sense. This was actually how I implemented it to start with, > but I changed it to be less intrusive on the nvme path. Let's revert the > patch and we can add it back when we push the 2.0 patches. > >> But before touching much of NVMe, I'd really, really like to see an >> actual spec first. > > The 2.0 spec. is open and is available here [1]. I thought you had > looked into it already... Anyway, feedback is more than welcome. > > [1] https://docs.google.com/document/d/1kedBY_1-hfkAlqT4EdwY6gz-6UOZbn7kIjWpmBLPNj0 > > Javier The 2.0 spec is still under development. No reason to redo the I/O stacks until it is final.
