On July 18, 2023 6:18:36 AM PDT, Zeng Guang <guang.zeng@xxxxxxxxx> wrote: >Linear Address Space Separation (LASS)[1] is a new mechanism that >enforces the same mode-based protections as paging, i.e. SMAP/SMEP >but without traversing the paging structures. Because the protections >enforced by LASS are applied before paging, "probes" by malicious >software will provide no paging-based timing information. > >Based on a linear-address organization, LASS partitions 64-bit linear >address space into two halves, user-mode address (LA[bit 63]=0) and >supervisor-mode address (LA[bit 63]=1). > >LASS aims to prevent any attempt to probe supervisor-mode addresses by >user mode, and likewise stop any attempt to access (if SMAP enabled) or >execute user-mode addresses from supervisor mode. > >When platform has LASS capability, KVM requires to expose this feature >to guest VM enumerated by CPUID.(EAX=07H.ECX=1):EAX.LASS[bit 6], and >allow guest to enable it via CR4.LASS[bit 27] on demand. For instruction >executed in the guest directly, hardware will perform the check. But KVM >also needs to behave same as hardware to apply LASS to kinds of guest >memory accesses when emulating instructions by software. > >KVM will take following LASS violations check on emulation path. >User-mode access to supervisor space address: > LA[bit 63] && (CPL == 3) >Supervisor-mode access to user space address: > Instruction fetch: !LA[bit 63] && (CPL < 3) > Data access: !LA[bit 63] && (CR4.SMAP==1) && ((RFLAGS.AC == 0 && > CPL < 3) || Implicit supervisor access) > >This patch series provide a LASS KVM solution and depends on kernel >enabling that can be found at >https://lore.kernel.org/all/20230609183632.48706-1-alexander.shishkin@xxxxxxxxxxxxxxx/ > >We tested the basic function of LASS virtualization including LASS >enumeration and enabling in non-root and nested environment. As KVM >unittest framework is not compatible to LASS rule, we use kernel module >and application test to emulate LASS violation instead. With KVM forced >emulation mechanism, we also verified the LASS functionality on some >emulation path with instruction fetch and data access to have same >behavior as hardware. > >How to extend kselftest to support LASS is under investigation and >experiment. > >[1] Intel ISE https://cdrdv2.intel.com/v1/dl/getContent/671368 >Chapter Linear Address Space Separation (LASS) > >------------------------------------------------------------------------ > >v1->v2 >1. refactor and optimize the interface of instruction emulation > by introducing new set of operation type definition prefixed with > "X86EMUL_F_" to distinguish access. >2. reorganize the patch to make each area of KVM better isolated. >3. refine LASS violation check design with consideration of wraparound > access across address space boundary. > >v0->v1 >1. Adapt to new __linearize() API >2. Function refactor of vmx_check_lass() >3. Refine commit message to be more precise >4. Drop LASS kvm cap detection depending > on hardware capability > >Binbin Wu (4): > KVM: x86: Consolidate flags for __linearize() > KVM: x86: Use a new flag for branch instructions > KVM: x86: Add an emulation flag for implicit system access > KVM: x86: Add X86EMUL_F_INVTLB and pass it in em_invlpg() > >Zeng Guang (4): > KVM: emulator: Add emulation of LASS violation checks on linear > address > KVM: VMX: Implement and apply vmx_is_lass_violation() for LASS > protection > KVM: x86: Virtualize CR4.LASS > KVM: x86: Advertise LASS CPUID to user space > > arch/x86/include/asm/kvm-x86-ops.h | 3 ++- > arch/x86/include/asm/kvm_host.h | 5 +++- > arch/x86/kvm/cpuid.c | 5 ++-- > arch/x86/kvm/emulate.c | 37 ++++++++++++++++++++--------- > arch/x86/kvm/kvm_emulate.h | 9 +++++++ > arch/x86/kvm/vmx/nested.c | 3 ++- > arch/x86/kvm/vmx/sgx.c | 4 ++++ > arch/x86/kvm/vmx/vmx.c | 38 ++++++++++++++++++++++++++++++ > arch/x86/kvm/vmx/vmx.h | 3 +++ > arch/x86/kvm/x86.c | 10 ++++++++ > arch/x86/kvm/x86.h | 2 ++ > 11 files changed, 102 insertions(+), 17 deletions(-) > Equating this with SMEP/SMAP is backwards. LASS is something completely different: it makes it so *user space accesses* cannot even walk the kernel page tables (specifically, the negative half of the linear address space.) Such an access with immediately #PF: it is similar to always having U=0 in the uppermost level of the page tables, except with LASS enabled the CPU will not even touch the page tables in memory.
