On 19/07/19 10:59, Wanpeng Li wrote: > https://lkml.org/lkml/2017/11/14/891, "The scheduler will save the > guest fpu context when a vCPU thread is preempted, and restore it when > it is scheduled back in." But I can't find any scheduler codes do > this. That's because applying commit 240c35a37 was completely wrong. The idea before commit 240c35a37 was that you have the following FPU states: userspace (QEMU) guest --------------------------------------------------------------------------- processor vcpu->arch.guest_fpu >>> KVM_RUN: kvm_load_guest_fpu vcpu->arch.user_fpu processor >>> preempt out vcpu->arch.user_fpu current->thread.fpu >>> preempt in vcpu->arch.user_fpu processor >>> back to userspace >>> kvm_put_guest_fpu processor vcpu->arch.guest_fpu --------------------------------------------------------------------------- After removing user_fpu, QEMU's FPU state is destroyed when KVM_RUN is preempted. So that's already messed up (I'll send a revert), and given the diagram above your patch makes total sense. With the new lazy model we want to hook into kvm_vcpu_arch_load and get the state back to the processor from current->thread.fpu, and indeed switch_fpu_return is essentially copy_kernel_to_fpregs(¤t->thread. fpu->state). However I would keep the fpregs_assert_state_consistent in kvm_arch_vcpu_load, and also WARN_ON_ONCE(test_thread_flag(TIF_NEED_FPU_LOAD)) in vcpu_enter_guest. Paolo
