Messages in this thread | | | Subject | Re: [PATCH v13 8/8] x86/vsyscall/64: Fixup Shadow Stack and Indirect Branch Tracking for vsyscall emulation | From | "Yu, Yu-cheng" <> | Date | Wed, 30 Sep 2020 15:33:04 -0700 |
| |
On 9/29/2020 1:00 PM, Andy Lutomirski wrote: > On Tue, Sep 29, 2020 at 12:57 PM Andy Lutomirski <luto@kernel.org> wrote: >> >> On Tue, Sep 29, 2020 at 11:37 AM Yu, Yu-cheng <yu-cheng.yu@intel.com> wrote: >>> >>> On 9/28/2020 10:37 AM, Andy Lutomirski wrote: >>>> On Mon, Sep 28, 2020 at 9:59 AM Yu-cheng Yu <yu-cheng.yu@intel.com> wrote: >>>>> >>>>> On Fri, 2020-09-25 at 09:51 -0700, Andy Lutomirski wrote: >>>>>>> On Sep 25, 2020, at 9:48 AM, Yu, Yu-cheng <yu-cheng.yu@intel.com> wrote: >>>>> + >>>>> + cet = get_xsave_addr(&fpu->state.xsave, XFEATURE_CET_USER); >>>>> + if (!cet) { >>>>> + /* >>>>> + * This is an unlikely case where the task is >>>>> + * CET-enabled, but CET xstate is in INIT. >>>>> + */ >>>>> + WARN_ONCE(1, "CET is enabled, but no xstates"); >>>> >>>> "unlikely" doesn't really cover this. >>>> >>>>> + fpregs_unlock(); >>>>> + goto sigsegv; >>>>> + } >>>>> + >>>>> + if (cet->user_ssp && ((cet->user_ssp + 8) < TASK_SIZE_MAX)) >>>>> + cet->user_ssp += 8; >>>> >>>> This looks buggy. The condition should be "if SHSTK is on, then add 8 >>>> to user_ssp". If the result is noncanonical, then some appropriate >>>> exception should be generated, probably by the FPU restore code -- see >>>> below. You should be checking the SHSTK_EN bit, not SSP. >>> >>> Updated. Is this OK? I will resend the whole series later. >>> >>> Thanks, >>> Yu-cheng >>> >>> ====== >>> >>> From 09803e66dca38d7784e32687d0693550948199ed Mon Sep 17 00:00:00 2001 >>> From: Yu-cheng Yu <yu-cheng.yu@intel.com> >>> Date: Thu, 29 Nov 2018 14:15:38 -0800 >>> Subject: [PATCH v13 8/8] x86/vsyscall/64: Fixup Shadow Stack and >>> Indirect Branch >>> Tracking for vsyscall emulation >>> >>> Vsyscall entry points are effectively branch targets. Mark them with >>> ENDBR64 opcodes. When emulating the RET instruction, unwind shadow stack >>> and reset IBT state machine. >>> >>> Signed-off-by: Yu-cheng Yu <yu-cheng.yu@intel.com> >>> --- >>> v13: >>> - Check shadow stack address is canonical. >>> - Change from writing to MSRs to writing to CET xstate. >>> >>> arch/x86/entry/vsyscall/vsyscall_64.c | 34 +++++++++++++++++++++++ >>> arch/x86/entry/vsyscall/vsyscall_emu_64.S | 9 ++++++ >>> arch/x86/entry/vsyscall/vsyscall_trace.h | 1 + >>> 3 files changed, 44 insertions(+) >>> >>> diff --git a/arch/x86/entry/vsyscall/vsyscall_64.c >>> b/arch/x86/entry/vsyscall/vsyscall_64.c >>> index 44c33103a955..30b166091d46 100644 >>> --- a/arch/x86/entry/vsyscall/vsyscall_64.c >>> +++ b/arch/x86/entry/vsyscall/vsyscall_64.c >>> @@ -38,6 +38,9 @@ >>> #include <asm/fixmap.h> >>> #include <asm/traps.h> >>> #include <asm/paravirt.h> >>> +#include <asm/fpu/xstate.h> >>> +#include <asm/fpu/types.h> >>> +#include <asm/fpu/internal.h> >>> >>> #define CREATE_TRACE_POINTS >>> #include "vsyscall_trace.h" >>> @@ -286,6 +289,44 @@ bool emulate_vsyscall(unsigned long error_code, >>> /* Emulate a ret instruction. */ >>> regs->ip = caller; >>> regs->sp += 8; >>> + >>> +#ifdef CONFIG_X86_CET >>> + if (tsk->thread.cet.shstk_size || tsk->thread.cet.ibt_enabled) { >>> + struct cet_user_state *cet; >>> + struct fpu *fpu; >>> + >>> + fpu = &tsk->thread.fpu; >>> + fpregs_lock(); >>> + >>> + if (!test_thread_flag(TIF_NEED_FPU_LOAD)) { >>> + copy_fpregs_to_fpstate(fpu); >>> + set_thread_flag(TIF_NEED_FPU_LOAD); >>> + } >>> + >>> + cet = get_xsave_addr(&fpu->state.xsave, XFEATURE_CET_USER); >>> + if (!cet) { >>> + /* >>> + * This should not happen. The task is >>> + * CET-enabled, but CET xstate is in INIT. >>> + */ >> >> Can the comment explain better, please? I would say something like: >> >> If the kernel thinks this task has CET enabled (because >> tsk->thread.cet has one of the features enabled), then the >> corresponding bits must also be set in the CET XSAVES region. If the >> CET XSAVES region is in the INIT state, then the kernel's concept of >> the task's CET state is corrupt. >> >>> + WARN_ONCE(1, "CET is enabled, but no xstates"); >>> + fpregs_unlock(); >>> + goto sigsegv; >>> + } >>> + >>> + if (cet->user_cet & CET_SHSTK_EN) { >>> + if (cet->user_ssp && (cet->user_ssp + 8 < TASK_SIZE_MAX)) >>> + cet->user_ssp += 8; >>> + } >> >> This makes so sense to me. Also, the vsyscall emulation code is >> intended to be as rigid as possible to minimize the chance that it >> gets used as an exploit gadget. So we should not silently corrupt >> anything. Moreover, this code seems quite dangerous -- you've created >> a gadget that does RET without actually verifying the SHSTK token. If >> SHSTK and some form of strong indirect branch/call CFI is in use, then >> the existance of a CFI-bypassing return primitive at a fixed address >> seems quite problematic. >> >> So I think you need to write a function that reasonably accurately >> emulates a usermode RET. >> > > For what it's worth, I think there is an alternative. If you all > (userspace people, etc) can come up with a credible way for a user > program to statically declare that it doesn't need vsyscalls, then we > could make SHSTK depend on *that*, and we could avoid this mess. This > breaks orthogonality, but it's probably a decent outcome. >
Would an arch_prctl(DISABLE_VSYSCALL) work? The kernel then sets a thread flag, and in emulate_vsyscall(), checks the flag.
When CET is enabled, ld-linux will do DISABLE_VSYSCALL.
How is that?
Yu-cheng
| |