On Fri, Dec 24, 2021, Chao Peng wrote: > On Thu, Dec 23, 2021 at 06:02:33PM +0000, Sean Christopherson wrote: > > On Thu, Dec 23, 2021, Chao Peng wrote: > > > Similar to hva_tree for hva range, maintain interval tree ofs_tree for > > > offset range of a fd-based memslot so the lookup by offset range can be > > > faster when memslot count is high. > > > > This won't work. The hva_tree relies on there being exactly one virtual address > > space, whereas with private memory, userspace can map multiple files into the > > guest at different gfns, but with overlapping offsets. > > OK, that's the point. > > > > > I also dislike hijacking __kvm_handle_hva_range() in patch 07. > > > > KVM also needs to disallow mapping the same file+offset into multiple gfns, which > > I don't see anywhere in this series. > > This can be checked against file+offset overlapping with existing slots > when register a new one. > > > > > In other words, there needs to be a 1:1 gfn:file+offset mapping. Since userspace > > likely wants to allocate a single file for guest private memory and map it into > > multiple discontiguous slots, e.g. to skip the PCI hole, the best idea off the top > > of my head would be to register the notifier on a per-slot basis, not a per-VM > > basis. It would require a 'struct kvm *' in 'struct kvm_memory_slot', but that's > > not a huge deal. > > > > That way, KVM's notifier callback already knows the memslot and can compute overlap > > between the memslot and the range by reversing the math done by kvm_memfd_get_pfn(). > > Then, armed with the gfn and slot, invalidation is just a matter of constructing > > a struct kvm_gfn_range and invoking kvm_unmap_gfn_range(). > > KVM is easy but the kernel bits would be difficulty, it has to maintain > fd+offset to memslot mapping because one fd can have multiple memslots, > it need decide which memslot needs to be notified. No, the kernel side maintains an opaque pointer like it does today, KVM handles reverse engineering the memslot to get the offset and whatever else it needs. notify_fallocate() and other callbacks are unchanged, though they probably can drop the inode. E.g. likely with bad math and handwaving on the overlap detection: int kvm_private_fd_fallocate_range(void *owner, pgoff_t start, pgoff_t end) { struct kvm_memory_slot *slot = owner; struct kvm_gfn_range gfn_range = { .slot = slot, .start = (start - slot->private_offset) >> PAGE_SHIFT, .end = (end - slot->private_offset) >> PAGE_SHIFT, .may_block = true, }; if (!has_overlap(slot, start, end)) return 0; gfn_range.end = min(gfn_range.end, slot->base_gfn + slot->npages); kvm_unmap_gfn_range(slot->kvm, &gfn_range); return 0; }