Re: [Qemu-devel] [PATCH v3 0/5] kvm "virtio pmem" device

[Pankaj Gupta <pagupta-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

> > > > > Right. Thinking about this I would be more concerned about the fact
> > > > > that
> > > > > guest can effectively pin amount of host's page cache upto size of
> > > > > the
> > > > > device/file passed to guest as PMEM, can't it Pankaj? Or is there
> > > > > some
> > > > > QEMU
> > > > > magic that avoids this?
> > > >
> > > > Yes, guest will pin these host page cache pages using 'get_user_pages'
> > > > by
> > > > elevating the page reference count. But these pages can be reclaimed by
> > > > host
> > > > at any time when there is memory pressure.
> > > 
> > > Wait, how can the guest pin the host pages? I would expect this to
> > > happen only when using vfio and device assignment. Otherwise, no the
> > > host can't reclaim a pinned page, that's the whole point of a pin to
> > > prevent the mm from reclaiming ownership.
> > 
> > yes. You are right I just used the pin word but it does not actually pin
> > pages
> > permanently. I had gone through the discussion on existing problems with
> > get_user_pages and DMA e.g [1] to understand Jan's POV. It does mention GUP
> > pin pages so I also used the word 'pin'. But guest does not permanently pin
> > these pages and these pages can be reclaimed by host.
> 
> OK, then I was just confused how virtio-pmem is going to work. Thanks for
> explanation! So can I imagine this as guest mmaping the host file and
> providing the mapped range as "NVDIMM pages" to the kernel inside the
> guest? Or is it more complex?

yes, that's correct. Host's Qemu process virtual address range is used as guest physical 
address and a direct mapping(EPT/NPT) is established. At guest side, this physical memory
range is plugged into guest system memory map and DAX mapping is setup using nvdimm calls.

Thanks,
Pankaj