On Mon, 25 Nov 2013 14:35:18 +0100 Paolo Bonzini <pbonz...@redhat.com> wrote:
> Il 25/11/2013 12:39, Gerd Hoffmann ha scritto: > >> > As use case 32-bit guest could start whit small initial memory > >> > and hotplug additional memory if needed up to point where 32-bit > >> > PCI hole starts. That would allow guests to launch with small amount > >> > but baloon up upto 2-3.5 Gb depending on machine type. > >> > I could drop 32-bit guest support and do only high mem hotplug if > >> > this case it not interesting to the comunity, any suggestions? > > 32bit limits start to hurt with 1GB already. Kernel address space is 1G > > on 32bit, so the kernel can't map all RAM all the time any more. Which > > in turn adds overhead for mapping/unmapping pages if the kernel must > > access highmem pages. So it's better to run 64bit guests even with alot > > less than 4G of memory. > > > > I'd tend to just not support 32bit guests, I think it simply isn't worth > > the trouble. > > Also because it's just non-PAE 32-bit guests, no? PAE guests would > support hotplug just fine. Yes, it shouldn't be issue for PAE in general. but it might not work for Windows Server 2003 (including 64-bit one), since it BSODed when it sees 64bit CRS (we saw it playing with 64-bit PCI window). Not sure if we care about it though. > Paolo >