Re: [PATCH 1/4] x86/percpu: Use explicit segment registers in lib/cmpxchg{8,16}b_emu.S

[Brian Gerst <brgerst@xxxxxxxxx>]

On Thu, Oct 12, 2023 at 12:13 PM Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>
> PER_CPU_VAR macro is intended to be applied to a symbol, it is not
> intended to be used as a selector between %fs and %gs segment
> registers for general operands.
>
> The address to these emulation functions is passed in a register, so
> use explicit segment registers to access percpu variable instead.
>
> Also add a missing function comment to this_cpu_cmpxchg8b_emu.
>
> No functional changes intended.
>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Cc: Borislav Petkov <bp@xxxxxxxxx>
> Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> ---
>  arch/x86/lib/cmpxchg16b_emu.S | 12 ++++++------
>  arch/x86/lib/cmpxchg8b_emu.S  | 30 +++++++++++++++++++++---------
>  2 files changed, 27 insertions(+), 15 deletions(-)
>
> diff --git a/arch/x86/lib/cmpxchg16b_emu.S b/arch/x86/lib/cmpxchg16b_emu.S
> index 6962df315793..2bd8b89bce75 100644
> --- a/arch/x86/lib/cmpxchg16b_emu.S
> +++ b/arch/x86/lib/cmpxchg16b_emu.S
> @@ -23,14 +23,14 @@ SYM_FUNC_START(this_cpu_cmpxchg16b_emu)
>         cli
>
>         /* if (*ptr == old) */
> -       cmpq    PER_CPU_VAR(0(%rsi)), %rax
> +       cmpq    %gs:(%rsi), %rax
>         jne     .Lnot_same
> -       cmpq    PER_CPU_VAR(8(%rsi)), %rdx
> +       cmpq    %gs:8(%rsi), %rdx
>         jne     .Lnot_same
>
>         /* *ptr = new */
> -       movq    %rbx, PER_CPU_VAR(0(%rsi))
> -       movq    %rcx, PER_CPU_VAR(8(%rsi))
> +       movq    %rbx, %gs:(%rsi)
> +       movq    %rcx, %gs:8(%rsi)
>
>         /* set ZF in EFLAGS to indicate success */
>         orl     $X86_EFLAGS_ZF, (%rsp)
> @@ -42,8 +42,8 @@ SYM_FUNC_START(this_cpu_cmpxchg16b_emu)
>         /* *ptr != old */
>
>         /* old = *ptr */
> -       movq    PER_CPU_VAR(0(%rsi)), %rax
> -       movq    PER_CPU_VAR(8(%rsi)), %rdx
> +       movq    %gs:(%rsi), %rax
> +       movq    %gs:8(%rsi), %rdx
>
>         /* clear ZF in EFLAGS to indicate failure */
>         andl    $(~X86_EFLAGS_ZF), (%rsp)
> diff --git a/arch/x86/lib/cmpxchg8b_emu.S b/arch/x86/lib/cmpxchg8b_emu.S
> index 49805257b125..b7d68d5e2d31 100644
> --- a/arch/x86/lib/cmpxchg8b_emu.S
> +++ b/arch/x86/lib/cmpxchg8b_emu.S
> @@ -24,12 +24,12 @@ SYM_FUNC_START(cmpxchg8b_emu)
>         pushfl
>         cli
>
> -       cmpl    0(%esi), %eax
> +       cmpl    (%esi), %eax
>         jne     .Lnot_same
>         cmpl    4(%esi), %edx
>         jne     .Lnot_same
>
> -       movl    %ebx, 0(%esi)
> +       movl    %ebx, (%esi)
>         movl    %ecx, 4(%esi)
>
>         orl     $X86_EFLAGS_ZF, (%esp)
> @@ -38,7 +38,7 @@ SYM_FUNC_START(cmpxchg8b_emu)
>         RET
>
>  .Lnot_same:
> -       movl    0(%esi), %eax
> +       movl    (%esi), %eax
>         movl    4(%esi), %edx
>
>         andl    $(~X86_EFLAGS_ZF), (%esp)
> @@ -53,18 +53,30 @@ EXPORT_SYMBOL(cmpxchg8b_emu)
>
>  #ifndef CONFIG_UML
>
> +/*
> + * Emulate 'cmpxchg8b %fs:(%esi)'
> + *
> + * Inputs:
> + * %esi : memory location to compare
> + * %eax : low 32 bits of old value
> + * %edx : high 32 bits of old value
> + * %ebx : low 32 bits of new value
> + * %ecx : high 32 bits of new value
> + *
> + * Notably this is not LOCK prefixed and is not safe against NMIs
> + */
>  SYM_FUNC_START(this_cpu_cmpxchg8b_emu)
>
>         pushfl
>         cli
>
> -       cmpl    PER_CPU_VAR(0(%esi)), %eax
> +       cmpl    %fs:(%esi), %eax
>         jne     .Lnot_same2
> -       cmpl    PER_CPU_VAR(4(%esi)), %edx
> +       cmpl    %fs:4(%esi), %edx
>         jne     .Lnot_same2
>
> -       movl    %ebx, PER_CPU_VAR(0(%esi))
> -       movl    %ecx, PER_CPU_VAR(4(%esi))
> +       movl    %ebx, %fs:(%esi)
> +       movl    %ecx, %fs:4(%esi)
>
>         orl     $X86_EFLAGS_ZF, (%esp)
>
> @@ -72,8 +84,8 @@ SYM_FUNC_START(this_cpu_cmpxchg8b_emu)
>         RET
>
>  .Lnot_same2:
> -       movl    PER_CPU_VAR(0(%esi)), %eax
> -       movl    PER_CPU_VAR(4(%esi)), %edx
> +       movl    %fs:(%esi), %eax
> +       movl    %fs:4(%esi), %edx
>
>         andl    $(~X86_EFLAGS_ZF), (%esp)
>
> --
> 2.41.0
>

This will break on !SMP builds, where per-cpu variables are just
regular data and not accessed with a segment prefix.

Brian Gerst