Michael, You once advocated idea of using MAP_NORESERVE to reserve continuous HVA upto maxmmem in QEMU and then 'allocating' guest's RAM from that range so that translation map for vhost could consist only from that single pre-reserved HVA range and and if guest access page outside of actually present memory it would be ok to let guest misbehave.
Reason we get so many fragments is holes pocked in initial RAM by device memory/MMIO ranges. Maybe we can reuse 'would be ok to let guest misbehave' part in vhost_set_memory() which tracks flat memory map represented by sections. Idea here is that each section has reference to a MemoryRegion it belongs to, for vhost mem map we could reuse that MemoryRegion range instead of set of sections that belong to it. Well behaved guest will continue working as it uses accessible RAM pages and malicious guest will misbehave if it will ask for translation of the page outside of accessible RAM. This way vhost mem map will typically have 1/2 entries for low/high mem ranges + an entry per each dimm. This way we can keep the same or less number of entries in vhost map without merging (as in practice it works only for fragmented initial memory and by chanse may work for dimm device) and at the same time number of entries will be constant (number of RAM memory regions) depending on initial RAM amount and amount of used dimm devices regardless of the order they were created or the runtime point they are created at.
