On 2020/12/14 23:36, Peter Xu wrote: > On Mon, Dec 14, 2020 at 10:14:11AM +0800, zhukeqian wrote: > > [...] > >>>>> Though indeed I must confess I don't know how it worked in general when >>>>> host >>>>> page size != target page size, at least for migration. For example, I >>>>> believe >>>>> kvm dirty logging is host page size based, though migration should be >>>>> migrating >>>>> pages in guest page size granule when it spots a dirty bit set. > > [1] > >> Hi Peter, > > Keqian, > >>> OTOH I'm more worried on the other question on how we handle guest psize != >>> host psize case for migration now... >> I think it does not matter when guest_psize != host_psize, as we only need >> to interact with >> stage2 page tables during migration. Stage2 is enough to tracking guest >> dirty memory, and even >> if guest close stage1, we also can do a successful migration. > > I don't know why 2-stage matters here, since I believe KVM can track dirty > pages either using two dimentional paging or shadowing, however it's always > done in host small page size. The question I'm confused is there seems to > have > a size mismatch between qemu migration and what kvm does [1]. For example, > how > migration works on ARM64 where host has psize==4K while guest has psize=64K. > Hi Peter,
OK, I got it ;-) Do you mean qemu_real_host_page_size != TARGET_PAGE_SIZE? After my analysis, I see that when qemu_real_host_page_size != TARGET_PAGE_SIZE, there are some problems indeed. I have send out some patches, please check whether they solve this problem, thanks! Keqian. > Thanks, >
