Re: [RFC 2/2] KVM: add virtio-pmem driver
- Date: Thu, 12 Oct 2017 14:54:53 -0700
- From: Dan Williams <dan.j.williams@xxxxxxxxx>
- Subject: Re: [RFC 2/2] KVM: add virtio-pmem driver
On Thu, Oct 12, 2017 at 2:25 PM, Pankaj Gupta <pagupta@xxxxxxxxxx> wrote:
>> > This patch adds virtio-pmem driver for KVM guest.
>> > Guest reads the persistent memory range information
>> > over virtio bus from Qemu and reserves the range
>> > as persistent memory. Guest also allocates a block
>> > device corresponding to the pmem range which later
>> > can be accessed with DAX compatible file systems.
>> > Idea is to use the virtio channel between guest and
>> > host to perform the block device flush for guest pmem
>> > DAX device.
>> > There is work to do including DAX file system support
>> > and other advanced features.
>> > Signed-off-by: Pankaj Gupta <pagupta@xxxxxxxxxx>
>> > ---
>> > drivers/virtio/Kconfig | 10 ++
>> > drivers/virtio/Makefile | 1 +
>> > drivers/virtio/virtio_pmem.c | 322
>> > +++++++++++++++++++++++++++++++++++++++
>> > include/uapi/linux/virtio_pmem.h | 55 +++++++
>> > 4 files changed, 388 insertions(+)
>> > create mode 100644 drivers/virtio/virtio_pmem.c
>> > create mode 100644 include/uapi/linux/virtio_pmem.h
>> > diff --git a/drivers/virtio/Kconfig b/drivers/virtio/Kconfig
>> > index cff773f15b7e..0192c4bda54b 100644
>> > --- a/drivers/virtio/Kconfig
>> > +++ b/drivers/virtio/Kconfig
>> > @@ -38,6 +38,16 @@ config VIRTIO_PCI_LEGACY
>> > If unsure, say Y.
>> > +config VIRTIO_PMEM
>> > + tristate "Virtio pmem driver"
>> > + depends on VIRTIO
>> > + ---help---
>> > + This driver adds persistent memory range within a KVM guest.
>> I think we need to call this something other than persistent memory to
>> make it clear that this not memory where the persistence can be
>> managed from userspace. The persistence point always requires a driver
>> call, so this is something distinctly different than "persistent
>> memory". For example, it's a bug if this memory range ends up backing
>> a device-dax range in the guest where there is no such thing as a
>> driver callback to perform the flushing. How does this solution
>> protect against that scenario?
> yes, you are right we are not providing device_dax in this case so it should
> be clear from name. Any suggestion for name?
So currently /proc/iomem in a guest with a pmem device attached to a
namespace looks like this:
c00000000-13bfffffff : Persistent Memory
c00000000-13bfffffff : namespace2.0
Can we call it "Virtio Shared Memory" to make it clear it is a
different beast than typical "Persistent Memory"? You can likely
inject your own name into the resource tree the same way we do in the
NFIT driver. See acpi_nfit_insert_resource().