On Saturday, 2022-02-12 at 12:02:14 +01, Paolo Bonzini wrote: > On 2/11/22 07:07, Leonardo Bras wrote: >> During host/guest switch (like in kvm_arch_vcpu_ioctl_run()), the kernel >> swaps the fpu between host/guest contexts, by using fpu_swap_kvm_fpstate(). >> When xsave feature is available, the fpu swap is done by: >> - xsave(s) instruction, with guest's fpstate->xfeatures as mask, is used >> to store the current state of the fpu registers to a buffer. >> - xrstor(s) instruction, with (fpu_kernel_cfg.max_features & >> XFEATURE_MASK_FPSTATE) as mask, is used to put the buffer into fpu regs. >> For xsave(s) the mask is used to limit what parts of the fpu regs >> will >> be copied to the buffer. Likewise on xrstor(s), the mask is used to >> limit what parts of the fpu regs will be changed. >> The mask for xsave(s), the guest's fpstate->xfeatures, is defined on >> kvm_arch_vcpu_create(), which (in summary) sets it to all features >> supported by the cpu which are enabled on kernel config. >> This means that xsave(s) will save to guest buffer all the fpu regs >> contents the cpu has enabled when the guest is paused, even if they >> are not used. >> This would not be an issue, if xrstor(s) would also do that. >> xrstor(s)'s mask for host/guest swap is basically every valid >> feature >> contained in kernel config, except XFEATURE_MASK_PKRU. >> Accordingto kernel src, it is instead switched in switch_to() and >> flush_thread(). >> Then, the following happens with a host supporting PKRU starts a >> guest that does not support it: >> 1 - Host has XFEATURE_MASK_PKRU set. 1st switch to guest, >> 2 - xsave(s) fpu regs to host fpustate (buffer has XFEATURE_MASK_PKRU) >> 3 - xrstor(s) guest fpustate to fpu regs (fpu regs have XFEATURE_MASK_PKRU) >> 4 - guest runs, then switch back to host, >> 5 - xsave(s) fpu regs to guest fpstate (buffer now have XFEATURE_MASK_PKRU) >> 6 - xrstor(s) host fpstate to fpu regs. >> 7 - kvm_vcpu_ioctl_x86_get_xsave() copy guest fpstate to userspace (with >> XFEATURE_MASK_PKRU, which should not be supported by guest vcpu) >> On 5, even though the guest does not support PKRU, it does have the >> flag >> set on guest fpstate, which is transferred to userspace via vcpu ioctl >> KVM_GET_XSAVE. >> This becomes a problem when the user decides on migrating the above >> guest >> to another machine that does not support PKRU: >> The new host restores guest's fpu regs to as they were before (xrstor(s)), >> but since the new host don't support PKRU, a general-protection exception >> ocurs in xrstor(s) and that crashes the guest. >> This can be solved by making the guest's fpstate->user_xfeatures >> only hold >> values compatible to guest_supported_xcr0. This way, on 7 the only flags >> copied to userspace will be the ones compatible to guest requirements, >> and thus there will be no issue during migration. >> As a bonus, will also fail if userspace tries to set fpu features >> that are not compatible to the guest configuration. (KVM_SET_XSAVE ioctl) >> Signed-off-by: Leonardo Bras <leobras@xxxxxxxxxx> >> --- >> arch/x86/kernel/fpu/core.c | 1 + >> arch/x86/kvm/cpuid.c | 4 ++++ >> 2 files changed, 5 insertions(+) >> diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c >> index 8dea01ffc5c1..e83d8b1fbc83 100644 >> --- a/arch/x86/kernel/fpu/core.c >> +++ b/arch/x86/kernel/fpu/core.c >> @@ -34,6 +34,7 @@ DEFINE_PER_CPU(u64, xfd_state); >> /* The FPU state configuration data for kernel and user space */ >> struct fpu_state_config fpu_kernel_cfg __ro_after_init; >> struct fpu_state_config fpu_user_cfg __ro_after_init; >> +EXPORT_SYMBOL(fpu_user_cfg); >> /* >> * Represents the initial FPU state. It's mostly (but not completely) zeroes, >> diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c >> index 494d4d351859..aecebd6bc490 100644 >> --- a/arch/x86/kvm/cpuid.c >> +++ b/arch/x86/kvm/cpuid.c >> @@ -296,6 +296,10 @@ static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu) >> vcpu->arch.guest_supported_xcr0 = >> cpuid_get_supported_xcr0(vcpu->arch.cpuid_entries, vcpu->arch.cpuid_nent); >> + /* Mask out features unsupported by guest */ >> + vcpu->arch.guest_fpu.fpstate->user_xfeatures = >> + fpu_user_cfg.default_features & vcpu->arch.guest_supported_xcr0; > > This is not correct, because default_features does not include the > optional features (such as AMX) that were the original reason to > go through all this mess. What about: > > vcpu->arch.guest_fpu.fpstate->user_xfeatures = > vcpu->arch.guest_fpu.fpstate->xfeatures & vcpu->arch.guest_supported_xcr0; > > ? Sorry if this is a daft question: In what situations will there be bits set in vcpu->arch.guest_supported_xcr0 that are not set in vcpu->arch.guest_fpu.fpstate->xfeatures ? guest_supported_xcr0 is filtered based on supported_xcr0, which I would expect to weed out all bits that are not set in ->xfeatures. > Paolo > >> kvm_update_pv_runtime(vcpu); >> vcpu->arch.maxphyaddr = cpuid_query_maxphyaddr(vcpu); dme. -- I used to get mad at my school, the teachers who taught me weren't cool.