Currently the memory core requires 16 bytes per guest page (0.4% overhead). This patchset, building on the patches posted last week, reduces the overhead to practically nothing. Assuming a 4k page size, contiguous aligned 4k, 4M, and 4G regions each occupy just 2 bytes. Since most guest memory is physically contiguous, we can describe a guest with just a few dozen kilobytes.
Please review carefully, especially patch 13, and test carefully, especially things like rom/device. Also available in git://git.kernel.org/pub/scm/virt/kvm/qemu-kvm.git memory/core Avi Kivity (20): memory: allow MemoryListeners to observe a specific address space xen: ignore I/O memory regions memory: split memory listener for the two address spaces memory: support stateless memory listeners memory: change memory registration to rebuild the memory map on each change memory: remove first level of l1_phys_map memory: unify phys_map last level with intermediate levels memory: store MemoryRegionSection pointers in phys_map memory: compress phys_map node pointers to 16 bits memory: fix RAM subpages in newly initialized pages memory: unify the two branches of cpu_register_physical_memory_log() memory: move tlb flush to MemoryListener commit callback memory: make phys_page_find() return a MemoryRegionSection memory: give phys_page_find() its own tree search loop memory: simplify multipage/subpage registration memory: replace phys_page_find_alloc() with phys_page_set() memory: switch phys_page_set() to a recursive implementation memory: change phys_page_set() to set multiple pages memory: unify PhysPageEntry::node and ::leaf memory: allow phys_map tree paths to terminate early exec.c | 828 +++++++++++++++++++++++++++++++++++++---------------------- hw/vhost.c | 18 ++- kvm-all.c | 18 ++- memory.c | 50 ++++- memory.h | 7 +- xen-all.c | 18 ++- 6 files changed, 620 insertions(+), 319 deletions(-) -- 1.7.9