On Thu, Feb 18, 2021 at 04:00:05PM +0100, David Hildenbrand wrote:
> On 18.02.21 14:31, Michal Privoznik wrote:
> > The virtio-mem is paravirtualized mechanism of adding/removing
> > memory to/from a VM. A virtio-mem-pci device is split into blocks
> > of equal size which are then exposed (all or only a requested
> > portion of them) to the guest kernel to use as regular memory.
> > Therefore, the device has two important attributes:
> >
> > 1) block-size, which defines the size of a block
> > 2) requested-size, which defines how much memory (in bytes)
> > is the device requested to expose to the guest.
> >
> > The 'block-size' is configured on command line and immutable
> > throughout device's lifetime. The 'requested-size' can be set on
> > the command line too, but also is adjustable via monitor. In
> > fact, that is how management software places its requests to
> > change the memory allocation. If it wants to give more memory to
> > the guest it changes 'requested-size' to a bigger value, and if it
> > wants to shrink guest memory it changes the 'requested-size' to a
> > smaller value. Note, value of zero means that guest should
> > release all memory offered by the device. Of course, guest has to
> > cooperate. Therefore, there is a third attribute 'size' which is
> > read only and reflects how much memory the guest still has. This
> > can be different to 'requested-size', obviously. Because of name
> > clash, I've named it 'actualsize' and it is dealt with in future
> > commits (it is a runtime information anyway).
> >
> > In the backend, memory for virtio-mem is backed by usual objects:
> > memory-backend-{ram,file,memfd} and their size puts the cap on
> > the amount of memory that a virtio-mem device can offer to a
> > guest. But we are already able to express this info using <size/>
> > under <target/>.
> >
> > Therefore, we need only two more elements to cover 'block-size'
> > and 'requested-size' attributes. This is the XML I've came up
> > with:
> >
> > <memory model='virtio-mem'>
> > <source>
> > <nodemask>1-3</nodemask>
> > <pagesize unit='KiB'>2048</pagesize>
> > </source>
> > <target>
> > <size unit='KiB'>2097152</size>
> > <node>0</node>
> > <block unit='KiB'>2048</block>
> > <requested unit='KiB'>1048576</requested>
> > </target>
> > <address type='pci' domain='0x0000' bus='0x00' slot='0x04' function='0x0'/>
> > </memory>
> >
> > I hope by now it is obvious that:
> >
> > 1) 'requested-size' must be an integer multiple of
> > 'block-size', and
> > 2) virtio-mem-pci device goes onto PCI bus and thus needs PCI
> > address.
> >
> > Then there is a limitation that the minimal 'block-size' is
> > transparent huge page size (I'll leave this without explanation).
> >
> > Since now we have (possibly) two or more devices that allow
> > memory inflation/deflation and accounting for all of them (and
> > thus keeping <currentMemory/> updated) might be hard. Therefore,
> > I'm deliberately forbidding memballoon. It's okay - virtio-mem is
> > superior to memballoon anyway. We can always reevaluate later.
>
> That's a bad idea. It'll still be used for getting memory stats, free page
> hinting and free page reporting.
Yep, and this feature is broadly used in mgmt apps, so if we enforce this
mutual exclusion it is putting apps into a no-win scenario.
> Very weird use cases might even want to mix balloon inflation/deflation with
> virtio-mem ...
I'd feel less bad about blocking inflation/deflation, but that does go
against the general libvirt POV that we should leave usage policies upto
the mgmt app to decide upon.
Regards,
Daniel
--
|: https://berrange.com -o- https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org -o- https://fstop138.berrange.com :|
|: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
