Re: [Qemu-devel] [RFC 2/2] KVM: add virtio-pmem driver

[Stefan Hajnoczi]

On Wed, Oct 18, 2017 at 08:51:37AM -0700, Dan Williams wrote:
> On Wed, Oct 18, 2017 at 6:03 AM, Stefan Hajnoczi <address@hidden> wrote:
> > On Tue, Oct 17, 2017 at 04:30:41AM -0400, Pankaj Gupta wrote:
> >>
> >> > > Are you saying do it as existing i.e ACPI pmem like interface?
> >> > > The reason we have created this new driver is exiting pmem driver
> >> > > does not define proper semantics for guest flushing requests.
> >> >
> >> > At this point I'm caring about the Linux-internal interface, and
> >> > for that it should be integrated into the nvdimm subsystem and not
> >> > a block driver.  How the host <-> guest interface looks is a different
> >> > idea.
> >> >
> >> > >
> >> > > Regarding block support of driver, we want to achieve DAX support
> >> > > to bypass guest page cache. Also, we want to utilize existing DAX
> >> > > capable file-system interfaces(e.g fsync) from userspace file API's
> >> > > to trigger the host side flush request.
> >> >
> >> > Well, if you want to support XFS+DAX better don't make it a block
> >> > devices, because I'll post patches soon to stop using the block device
> >> > entirely for the DAX case.
> >>
> >> o.k I will look at your patches once they are in mailing list.
> >> Thanks for the heads up.
> >>
> >> If I am guessing it right, we don't need block device additional features
> >> for pmem? We can bypass block device features like blk device cache flush 
> >> etc.
> >> Also, still we would be supporting ext4 & XFS filesystem with pmem?
> >>
> >> If there is time to your patches can you please elaborate on this a bit.
> >
> > I think the idea is that the nvdimm subsystem already adds block device
> > semantics on top of the struct nvdimms that it manages.  See
> > drivers/nvdimm/blk.c.
> >
> > So it would be cleaner to make virtio-pmem an nvdimm bus.  This will
> > eliminate the duplication between your driver and drivers/nvdimm/ code.
> > Try "git grep nvdimm_bus_register" to find drivers that use the nvdimm
> > subsystem.
> 
> This use case is not "Persistent Memory". Persistent Memory is
> something you can map and make persistent with CPU instructions.
> Anything that requires a driver call is device driver managed "Shared
> Memory".

Dan, in that case do you have ideas regarding Christoph Hellwig's
comment that this driver should be integrated into the nvdimm subsystem
instead of a new block driver?

Stefan