On Mon, Jul 14, 2025 at 4:10 PM Ira Weiny <ira.weiny@xxxxxxxxx> wrote: > > Vishal Annapurve wrote: > > On Fri, Jul 11, 2025 at 3:56 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > > > > On Fri, Jul 11, 2025, Ira Weiny wrote: > > > > Michael Roth wrote: > > > > > For in-place conversion: the idea is that userspace will convert > > > > > private->shared to update in-place, then immediately convert back > > > > > shared->private; > > > > > > > > Why convert from private to shared and back to private? Userspace which > > > > knows about mmap and supports it should create shared pages, mmap, write > > > > data, then convert to private. > > > > > > Dunno if there's a strong usecase for converting to shared *and* populating the > > > data, but I also don't know that it's worth going out of our way to prevent such > > > behavior, at least not without a strong reason to do so. E.g. if it allowed for > > > a cleaner implementation or better semantics, then by all means. But I don't > > > think that's true here? Though I haven't thought hard about this, so don't > > > quote me on that. :-) > > > > If this is a huge page backing, starting as shared will split all the > > pages to 4K granularity upon allocation. > > Why? What is the reason it needs to be split? I think you and Sean have similar questions. This init private-> shared-> fill -> private scheme is for userspace for the initial guest payload population. Another choice userspace has is to begin the whole file as shared -> fill -> only needed ranges to private. Regarding shared memory ranges for CC VMs, guest_memfd huge page support [1] simply works by splitting hugepages in 4K chunks for shared regions to allow core-mm to manage the pages without affecting rest of the private ranges within a hugepage. The need for splitting has been discussed in MM alignment calls and LPC 2024[2]. [1] https://lore.kernel.org/lkml/cover.1747264138.git.ackerleytng@xxxxxxxxxx/ [2] https://lpc.events/event/18/contributions/1764/
