On Mon, Jul 08, 2019 at 03:52:29PM -0700, Matt Fitzpatrick wrote: > Hey Klaus, > > Sorry for the late reply! I finally found this message amid the pile of > emails Qemu dumped on me. > > I don't know what the right answer is here... NVMe is designed in a way > where you *do* "carve up" the flash into logical groupings and the nvme > firmware decides on how that's done. Those logical groupings can be attached > to different controllers(which we don't have here yet?) after init, but > that's a problem for future us I guess?But that's all stuff you already > know. >
Yeah, I havn't started worrying about that ;) > The "nvme-nvm" solution might be the right approach, but I'm a bit hesitant > on the idea of growing tnvmcap... > > I can't think of any way to create namespaces on the fly and not have it use > some single existing block backend, unless we defined a range of block > images on qemu start and namespace create/attach only uses one image up to > and including it's max size per namespace? That might work, and I think > that's what you suggested (or at least is similar to), though it could be > pretty wasteful. It wouldn't offer a "true" namespace management support, > but could be close enough. > Having an emulated device that supports namespace management would be very useful for testing software, but yeah, I have a hard time seeing how we can make that fit with the current "QEMU model". > I'm in the middle of going through the patch you posted. Nice job! I'm glad > to see more people adding enhancements. It was pretty stale for years. > Thanks for looking at it, I know it's a lot to go through ;) > -Matt > On 7/5/19 12:50 AM, Klaus Birkelund wrote: > > On Tue, Jul 02, 2019 at 10:39:36AM -0700, Matt Fitzpatrick wrote: > > > Adding namespace management support to the nvme device. Namespace creation > > > requires contiguous block space for a simple method of allocation. > > > > > > I wrote this a few years ago based on Keith's fork and nvmeqemu fork and > > > have recently re-synced with the latest trunk. Some data structures in > > > nvme.h are a bit more filled out that strictly necessary as this is also > > > the > > > base for sr-iov and IOD patched to be submitted later. > > > > > Hi Matt, > > > > Nice! I'm always happy when new features for the nvme device is posted! > > > > I'll be happy to review it, but I won't start going through it in > > details because I believe the approach to supporting multiple namespaces > > is flawed. We had a recent discussion on this and I also got some > > unrelated patches rejected due to implementing it similarly by carving > > up the image. > > > > I have posted a long series that includes a patch for multiple > > namespaces. It is implemented by introducing a fresh `nvme-ns` device > > model that represents a namespace and attaches to a bus created by the > > parent `nvme` controller device. > > > > The core issue is that a qemu image /should/ be attachable to other > > devices (say ide) and not strictly tied to the one device model. Thus, > > we cannot just shove a bunch of namespaces into a single image. > > > > But, in light of your patch, I'm not convinced that my implementation is > > the correct solution. Maybe the abstraction should not be an `nvme-ns` > > device, but a `nvme-nvm` device that when attached changes TNVMCAP and > > UNVMCAP? Maybe you have some input for this? Or we could have both and > > dynamically create the nvme-ns devices on top of nvme-nvm devices. I > > think it would still require a 1-to-1 mapping, but it could be a way to > > support the namespace management capability. > > > > > > Cheers, > > Klaus > > > Hi Kevin, This highlights another situation where the "1 image to 1 block device" model doesn't fit that well with NVMe. Especially with the introduction of "NVM Sets" in NVMe 1.4. It would be very nice to introduce a 'nvme-nvmset' device model that adds an NVM Set which the controller can then create namespaces in. Is it completely unacceptable for a device to use the image in such a way that it would not make sense (aka present the same block device) when attached to another device (ide, ...)? I really have a hard time seeing how we could support these features without violating the '1 image to 1 block device" model. Cheers, Klaus
