On Tue, Dec 10, 2013 at 11:02:41PM +0200, Michael S. Tsirkin wrote: > On Tue, Dec 10, 2013 at 03:21:44PM -0200, Marcelo Tosatti wrote: > > On Tue, Dec 10, 2013 at 01:05:42PM -0200, Marcelo Tosatti wrote: > > > On Tue, Dec 10, 2013 at 02:18:36PM +0100, Paolo Bonzini wrote: > > > > Il 28/11/2013 11:26, Michael S. Tsirkin ha scritto: > > > > > On Mon, Nov 25, 2013 at 06:43:13PM +0100, Paolo Bonzini wrote: > > > > >> v2: condition enablement of new mapping to new machine types (Paolo) > > > > >> v3: fix changelog > > > > >> v4: rebase > > > > >> v5: ensure alignment of piecetwo on 2MB GPA (Igor) > > > > >> do not register zero-sized piece-one (Igor) > > > > >> v6: fix memory leak (Igor) > > > > >> fix integer overflow (Igor) > > > > >> > > > > >> ---- > > > > >> > > > > >> Align guest physical address and host physical address > > > > >> beyond guest 4GB on a 1GB boundary. > > > > >> > > > > >> Otherwise 1GB TLBs cannot be cached for the range. > > > > >> > > > > >> Signed-off-by: Marcelo Tosatti <mtosa...@redhat.com> > > > > >> [Reorganize code, keep same logic. - Paolo] > > > > >> Signed-off-by: Paolo Bonzini <pbonz...@redhat.com> > > > > >> --- > > > > > > > > > > Applied, thanks. > > > > > > > > As discussed offlist, I'm not sure anymore that this is the right > > > > approach to the problem. No doubt it is very clever, in that it is > > > > absolutely transparent to the guest. However, the non-contiguous > > > > mapping of ram_addr_t makes it more complex to associate the right NUMA > > > > policy to the ranges. > > > > > > Please explain what is the difference, and why the complexity does not > > > exist with non-contiguous mapping of ram_addr_t. > > > > You are right - it forces the 1GB page which contains the hole > > to be on the same NUMA node as the tail 1GB page - otherwise > > incorrect NUMA assignment is not possible. > > What does this phrase mean? > > Are we all in agreement that we want this patch, in addition to > resizing below 4g memory?
It means that its necessary to expose that 3-4GB physical memory region in QEMU belongs to the same node (that is, guest must be aware that 3-3.75GB and the tail of RAM are on the same node). So the problem Paolo mentions is fixable.