Re: [PATCH v9 14/19] KVM: selftests: Add helpers to init TDX memory and finalize VM

[Binbin Wu <binbin.wu@linux.intel.com>]

On 8/26/2025 9:07 AM, Yan Zhao wrote:
> On Mon, Aug 25, 2025 at 02:02:00PM -0500, Sagi Shahar wrote:
>> On Mon, Aug 25, 2025 at 3:41 AM Yan Zhao <yan.y.zhao@intel.com> wrote:
>>> On Wed, Aug 20, 2025 at 09:29:07PM -0700, Sagi Shahar wrote:
>>>> From: Ackerley Tng <ackerleytng@google.com>
>>>>
>>>> TDX protected memory needs to be measured and encrypted before it can be
>>>> used by the guest. Traverse the VM's memory regions and initialize all
>>>> the protected ranges by calling KVM_TDX_INIT_MEM_REGION.
>>>>
>>>> Once all the memory is initialized, the VM can be finalized by calling
>>>> KVM_TDX_FINALIZE_VM.
>>>>
>>>> Signed-off-by: Ackerley Tng <ackerleytng@google.com>
>>>> Co-developed-by: Erdem Aktas <erdemaktas@google.com>
>>>> Signed-off-by: Erdem Aktas <erdemaktas@google.com>
>>>> Co-developed-by: Sagi Shahar <sagis@google.com>
>>>> Signed-off-by: Sagi Shahar <sagis@google.com>
>>>> ---
>>>>   .../selftests/kvm/include/x86/tdx/tdx_util.h  |  2 +
>>>>   .../selftests/kvm/lib/x86/tdx/tdx_util.c      | 97 +++++++++++++++++++
>>>>   2 files changed, 99 insertions(+)
>>>>
>>>> diff --git a/tools/testing/selftests/kvm/include/x86/tdx/tdx_util.h b/tools/testing/selftests/kvm/include/x86/tdx/tdx_util.h
>>>> index a2509959c7ce..2467b6c35557 100644
>>>> --- a/tools/testing/selftests/kvm/include/x86/tdx/tdx_util.h
>>>> +++ b/tools/testing/selftests/kvm/include/x86/tdx/tdx_util.h
>>>> @@ -71,4 +71,6 @@ void vm_tdx_load_common_boot_parameters(struct kvm_vm *vm);
>>>>   void vm_tdx_load_vcpu_boot_parameters(struct kvm_vm *vm, struct kvm_vcpu *vcpu);
>>>>   void vm_tdx_set_vcpu_entry_point(struct kvm_vcpu *vcpu, void *guest_code);
>>>>
>>>> +void vm_tdx_finalize(struct kvm_vm *vm);
>>>> +
>>>>   #endif // SELFTESTS_TDX_TDX_UTIL_H
>>>> diff --git a/tools/testing/selftests/kvm/lib/x86/tdx/tdx_util.c b/tools/testing/selftests/kvm/lib/x86/tdx/tdx_util.c
>>>> index d8eab99d9333..4024587ed3c2 100644
>>>> --- a/tools/testing/selftests/kvm/lib/x86/tdx/tdx_util.c
>>>> +++ b/tools/testing/selftests/kvm/lib/x86/tdx/tdx_util.c
>>>> @@ -274,3 +274,100 @@ void vm_tdx_init_vm(struct kvm_vm *vm, uint64_t attributes)
>>>>
>>>>        free(init_vm);
>>>>   }
>>>> +
>>>> +static void tdx_init_mem_region(struct kvm_vm *vm, void *source_pages,
>>>> +                             uint64_t gpa, uint64_t size)
>>>> +{
>>>> +     uint32_t metadata = KVM_TDX_MEASURE_MEMORY_REGION;
>>>> +     struct kvm_tdx_init_mem_region mem_region = {
>>>> +             .source_addr = (uint64_t)source_pages,
>>>> +             .gpa = gpa,
>>>> +             .nr_pages = size / PAGE_SIZE,
>>>> +     };
>>>> +     struct kvm_vcpu *vcpu;
>>>> +
>>>> +     vcpu = list_first_entry_or_null(&vm->vcpus, struct kvm_vcpu, list);
>>>> +
>>>> +     TEST_ASSERT((mem_region.nr_pages > 0) &&
>>>> +                 ((mem_region.nr_pages * PAGE_SIZE) == size),
>>>> +                 "Cannot add partial pages to the guest memory.\n");
>>>> +     TEST_ASSERT(((uint64_t)source_pages & (PAGE_SIZE - 1)) == 0,
>>>> +                 "Source memory buffer is not page aligned\n");
>>>> +     vm_tdx_vcpu_ioctl(vcpu, KVM_TDX_INIT_MEM_REGION, metadata, &mem_region);
>>>> +}
>>>> +
>>>> +static void tdx_init_pages(struct kvm_vm *vm, void *hva, uint64_t gpa,
>>>> +                        uint64_t size)
>>>> +{
>>>> +     void *scratch_page = calloc(1, PAGE_SIZE);
>>>> +     uint64_t nr_pages = size / PAGE_SIZE;
>>>> +     int i;
>>>> +
>>>> +     TEST_ASSERT(scratch_page,
>>>> +                 "Could not allocate memory for loading memory region");
>>>> +
>>>> +     for (i = 0; i < nr_pages; i++) {
>>>> +             memcpy(scratch_page, hva, PAGE_SIZE);
>>>> +
>>>> +             tdx_init_mem_region(vm, scratch_page, gpa, PAGE_SIZE);
>>>> +
>>>> +             hva += PAGE_SIZE;
>>>> +             gpa += PAGE_SIZE;
>>>> +     }
>>>> +
>>>> +     free(scratch_page);
>>>> +}
>>>> +
>>>> +static void load_td_private_memory(struct kvm_vm *vm)
>>>> +{
>>>> +     struct userspace_mem_region *region;
>>>> +     int ctr;
>>>> +
>>>> +     hash_for_each(vm->regions.slot_hash, ctr, region, slot_node) {
>>>> +             const struct sparsebit *protected_pages = region->protected_phy_pages;
>>>> +             const vm_paddr_t gpa_base = region->region.guest_phys_addr;
>>>> +             const uint64_t hva_base = region->region.userspace_addr;
>>>> +             const sparsebit_idx_t lowest_page_in_region = gpa_base >> vm->page_shift;
>>>> +
>>>> +             sparsebit_idx_t i;
>>>> +             sparsebit_idx_t j;
>>>> +
>>>> +             if (!sparsebit_any_set(protected_pages))
>>>> +                     continue;
>>>> +
>>>> +             sparsebit_for_each_set_range(protected_pages, i, j) {
>>>> +                     const uint64_t size_to_load = (j - i + 1) * vm->page_size;
>>>> +                     const uint64_t offset =
>>>> +                             (i - lowest_page_in_region) * vm->page_size;
>>>> +                     const uint64_t hva = hva_base + offset;
>>>> +                     const uint64_t gpa = gpa_base + offset;
>>>> +
>>>> +                     vm_set_memory_attributes(vm, gpa, size_to_load,
>>>> +                                              KVM_MEMORY_ATTRIBUTE_PRIVATE);
>>>> +
>>>> +                     /*
>>>> +                      * Here, memory is being loaded from hva to gpa. If the memory
>>>> +                      * mapped to hva is also used to back gpa, then a copy has to be
>>>> +                      * made just for loading, since KVM_TDX_INIT_MEM_REGION ioctl
>>>> +                      * cannot encrypt memory in place.
>>>> +                      *
>>>> +                      * To determine if memory mapped to hva is also used to back
>>>> +                      * gpa, use a heuristic:
>>>> +                      *
>>>> +                      * If this memslot has guest_memfd, then this memslot should
>>>> +                      * have memory backed from two sources: hva for shared memory
>>>> +                      * and gpa will be backed by guest_memfd.
>>>> +                      */
>>>> +                     if (region->region.guest_memfd == -1)
>>> Why to pass !guest_memfd region to tdx_init_mem_region()?
>>>
>> Not sure I understand your comment.
>  From the implementation of tdx_init_pages(), it also invokes
> tdx_init_mem_region(), which further invokes ioctl KVM_TDX_INIT_MEM_REGION.
>
> However, if the region is with guest_memfd == -1, the ioctl
> KVM_TDX_INIT_MEM_REGION should fail as kvm_gmem_populate() won't succeed.
>
> So, I'm wondering why there's a need to for the case of
> "region->region.guest_memfd == -1".
>
> Or anything I missed?
I had the same question in v8
https://lore.kernel.org/lkml/4b7e7099-79da-4178-8f16-6780d8137ae1@linux.intel.com/

I guess the code path for non-guest_memfd is due to some old versions of TDX KVM
code before upstream. Currently, KVM doesn't support private memory from
non-guest_memfd backed memory.

>
>>>> +                             tdx_init_pages(vm, (void *)hva, gpa, size_to_load);
>>>> +                     else
>>>> +                             tdx_init_mem_region(vm, (void *)hva, gpa, size_to_load);
>>>> +             }
>>>> +     }
>>>> +}
>>>> +
>>>> +void vm_tdx_finalize(struct kvm_vm *vm)
>>>> +{
>>>> +     load_td_private_memory(vm);
>>>> +     vm_tdx_vm_ioctl(vm, KVM_TDX_FINALIZE_VM, 0, NULL);
>>>> +}
>>>> --
>>>> 2.51.0.rc1.193.gad69d77794-goog
>>>>
>>>>