Re: [PATCH 2/6] x86/entry/64: Convert SYSRET validation tests to C

[Brian Gerst]

On Tue, Jul 18, 2023 at 11:21 AM Brian Gerst <brgerst@xxxxxxxxx> wrote:
>
> On Tue, Jul 18, 2023 at 10:49 AM Mika Penttilä <mpenttil@xxxxxxxxxx> wrote:
> >
> >
> >
> > On 18.7.2023 17.25, Brian Gerst wrote:
> > > On Tue, Jul 18, 2023 at 10:17 AM Mika Penttilä <mpenttil@xxxxxxxxxx> wrote:
> > >>
> > >> Hi,
> > >>
> > >>
> > >> On 18.7.2023 16.44, Brian Gerst wrote:
> > >>> Signed-off-by: Brian Gerst <brgerst@xxxxxxxxx>
> > >>> ---
> > >>>    arch/x86/entry/common.c        | 50 ++++++++++++++++++++++++++++++-
> > >>>    arch/x86/entry/entry_64.S      | 55 ++--------------------------------
> > >>>    arch/x86/include/asm/syscall.h |  2 +-
> > >>>    3 files changed, 52 insertions(+), 55 deletions(-)
> > >>>
> > >>> diff --git a/arch/x86/entry/common.c b/arch/x86/entry/common.c
> > >>> index 6c2826417b33..afe79c3f1c5b 100644
> > >>> --- a/arch/x86/entry/common.c
> > >>> +++ b/arch/x86/entry/common.c
> > >>> @@ -70,8 +70,12 @@ static __always_inline bool do_syscall_x32(struct pt_regs *regs, int nr)
> > >>>        return false;
> > >>>    }
> > >>>
> > >>> -__visible noinstr void do_syscall_64(struct pt_regs *regs, int nr)
> > >>> +/* Returns true to return using SYSRET, or false to use IRET */
> > >>> +__visible noinstr bool do_syscall_64(struct pt_regs *regs, int nr)
> > >>>    {
> > >>> +     long rip;
> > >>> +     unsigned int shift_rip;
> > >>> +
> > >>>        add_random_kstack_offset();
> > >>>        nr = syscall_enter_from_user_mode(regs, nr);
> > >>>
> > >>> @@ -84,6 +88,50 @@ __visible noinstr void do_syscall_64(struct pt_regs *regs, int nr)
> > >>>
> > >>>        instrumentation_end();
> > >>>        syscall_exit_to_user_mode(regs);
> > >>> +
> > >>> +     /*
> > >>> +      * Check that the register state is valid for using SYSRET to exit
> > >>> +      * to userspace.  Otherwise use the slower but fully capable IRET
> > >>> +      * exit path.
> > >>> +      */
> > >>> +
> > >>> +     /* XEN PV guests always use IRET path */
> > >>> +     if (cpu_feature_enabled(X86_FEATURE_XENPV))
> > >>> +             return false;
> > >>> +
> > >>> +     /* SYSRET requires RCX == RIP and R11 == EFLAGS */
> > >>> +     if (unlikely(regs->cx != regs->ip || regs->r11 != regs->flags))
> > >>> +             return false;
> > >>> +
> > >>> +     /* CS and SS must match the values set in MSR_STAR */
> > >>> +     if (unlikely(regs->cs != __USER_CS || regs->ss != __USER_DS))
> > >>> +             return false;
> > >>> +
> > >>> +     /*
> > >>> +      * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
> > >>> +      * in kernel space.  This essentially lets the user take over
> > >>> +      * the kernel, since userspace controls RSP.
> > >>> +      *
> > >>> +      * Change top bits to match most significant bit (47th or 56th bit
> > >>> +      * depending on paging mode) in the address.
> > >>> +      */
> > >>> +     shift_rip = (64 - __VIRTUAL_MASK_SHIFT + 1);
> > >>
> > >> Should this be:
> > >>
> > >>          shift_rip = (64 - __VIRTUAL_MASK_SHIFT - 1);
> > >> ?
> > >
> > > I removed a set of parentheses, which switched the sign from -1 to +1.
> > > I could put it back if that's less confusing.
> > >
> >
> > I mean isn't it supposed to be:
> > shift_rip = (64 - 48) for 4 level, now it's
> > shift_rip = (64 - 46)
> >
> > __VIRTUAL_MASK_SHIFT == 47

My apologies, you were right.  I've been sitting on this series for a
while and finally got around to posting it and didn't catch that
error.

>
> Original:
> (64 - (47 + 1)) = (64 - 48) = 16
>
>   c5:   48 c1 e1 10             shl    $0x10,%rcx
>   c9:   48 c1 f9 10             sar    $0x10,%rcx

This was wrong.  I hastily compiled this after I had reverted to the
original formula.

> New:
> (64 - 47 - 1) = (17 - 1) = 16
>
>  18b:   b9 10 00 00 00          mov    $0x10,%ecx
>  193:   48 d3 e2                shl    %cl,%rdx
>  196:   48 d3 fa                sar    %cl,%rdx
>
> Anyways, I'll switch it back to the original formula.  I'm not going
> to argue any more about basic math.

I'll send a v2 later after any more feedback.  Thanks.

 Brian Gerst