On Mon, 2025-04-28 at 18:11 -0700, Chang S. Bae wrote: > On 4/28/2025 8:42 AM, Edgecombe, Rick P wrote: > > > > Right, so there should be no need to keep a separate features and buffer size > > for KVM's xsave UABI, as this patch does. Let's just leave it using the core > > kernels UABI version. > > Hmm, why so? > > As I see it, the vcpu->arch.guest_fpu structure is already exposed to > KVM. This series doesn’t modify those structures (fpu_guest and > fpstate), other than removing a dead field (patch 2). > > Both ->usersize and ->user_xfeatures fields are already exposed -- > currently KVM just doesn’t reference them at all. > > All the changes introduced here are transparent to KVM. Organizing the > initial values and wiring up guest_perm and fpstate are entirely > internal to the x86 core, no? This patch adds struct vcpu_fpu_config, with new fields user_size, user_features. Then those fields are used to configure the guest FPU, where today it just uses fpu_user_cfg.default_features, etc. KVM doesn't refer to any of those fields specifically, but since they are used to configure struct fpu_guest they become part of KVM's uABI. Per Sean, KVM's KVM_GET_XSAVE API won't differ from arch/x86's uABI behavior. There is (and will be) only one default user features and size. So what is the point of having a special guest version with identical values? Just use the single one for guest FPU and normal. Chao mentioned offline it was for symmetry. I don't want to make a big deal out of it, but it doesn't make sense to me. It made me wonder if there was some divergence in KVM and arch/x86 user features.
