A virtio-mem device manages a memory region in guest physical address
space, represented as a single (currently large) memory region in QEMU,
mapped into system memory address space. Before the guest is allowed to use
memory blocks, it must coordinate with the hypervisor (plug blocks). After
a reboot, all memory is usually unplugged - when the guest comes up, it
detects the virtio-mem device and selects memory blocks to plug (based on
resize requests from the hypervisor).
Memory hot(un)plug consists of (un)plugging memory blocks via a virtio-mem
device (triggered by the guest). When unplugging blocks, we discard the
memory - similar to memory balloon inflation. In contrast to memory
ballooning, we always know which memory blocks a guest may actually use -
especially during a reboot, after a crash, or after kexec (and during
hibernation as well). Guests agreed to not access unplugged memory again,
especially not via DMA.
The issue with vfio is, that it cannot deal with random discards - for this
reason, virtio-mem and vfio can currently only run mutually exclusive.
Especially, vfio would currently map the whole memory region (with possible
only little/no plugged blocks), resulting in all pages getting pinned and
therefore resulting in a higher memory consumption than expected (turning
virtio-mem basically useless in these environments).
To make vfio work nicely with virtio-mem, we have to map only the plugged
blocks, and map/unmap properly when plugging/unplugging blocks (including
discarding of RAM when unplugging). We achieve that by using a new notifier
mechanism that communicates changes.