On Thu, Jan 09, 2020 at 02:08:52PM -0500, Michael S. Tsirkin wrote: > On Thu, Jan 09, 2020 at 12:08:49PM -0500, Peter Xu wrote: > > On Thu, Jan 09, 2020 at 11:40:23AM -0500, Michael S. Tsirkin wrote: > > > > [...] > > > > > > > I know it's mostly relevant for huge VMs, but OTOH these > > > > > probably use huge pages. > > > > > > > > Yes huge VMs could benefit more, especially if the dirty rate is not > > > > that high, I believe. Though, could you elaborate on why huge pages > > > > are special here? > > > > > > > > Thanks, > > > > > > With hugetlbfs there are less bits to test: e.g. with 2M pages a single > > > bit set marks 512 pages as dirty. We do not take advantage of this > > > but it looks like a rather obvious optimization. > > > > Right, but isn't that the trade-off between granularity of dirty > > tracking and how easy it is to collect the dirty bits? Say, it'll be > > merely impossible to migrate 1G-huge-page-backed guests if we track > > dirty bits using huge page granularity, since each touch of guest > > memory will cause another 1G memory to be transferred even if most of > > the content is the same. 2M can be somewhere in the middle, but still > > the same write amplify issue exists. > > > > OK I see I'm unclear. > > IIUC at the moment KVM never uses huge pages if any part of the huge page is > tracked. To be more precise - I think it's per-memslot. Say, if the memslot is dirty tracked, then no huge page on the host on that memslot (even if guest used huge page over that). > But if all parts of the page are written to then huge page > is used. I'm not sure of this... I think it's still in 4K granularity. > > In this situation the whole huge page is dirty and needs to be migrated. Note that in QEMU we always migrate pages in 4K for x86, iiuc (please refer to ram_save_host_page() in QEMU). > > > PS. that seems to be another topic after all besides the dirty ring > > series because we need to change our policy first if we want to track > > it with huge pages; with that, for dirty ring we can start to leverage > > the kvm_dirty_gfn.pad to store the page size with another new kvm cap > > when we really want. > > > > Thanks, > > Seems like leaking implementation detail to UAPI to me. I'd say it's not the only place we have an assumption at least (please also refer to uffd_msg.pagefault.address). IMHO it's not something wrong because interfaces can be extended, but I am open to extending kvm_dirty_gfn to cover a length/size or make the pad larger (as long as Paolo is fine with this). Thanks, -- Peter Xu
