Re: [PATCH 2/4] x86: eliminate most XPTI entry/exit code when it's not in use

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 23/01/18 10:37, Jan Beulich wrote:
> Introduce a synthetic feature flag to use alternative instruction
> patching to NOP out all code on entry/exit paths other than those
> involved in NMI/#MC handling (the patching logic can't properly handle
> those paths yet). Having NOPs here is generally better than using
> conditional branches.

Given my other series, I'd prefer to fix the IST paths rather than
introduce yet-more workarounds.

> --- a/xen/arch/x86/x86_64/compat/entry.S
> +++ b/xen/arch/x86/x86_64/compat/entry.S
> @@ -189,7 +189,7 @@ ENTRY(compat_post_handle_exception)
>  
>  /* See lstar_enter for entry register state. */
>  ENTRY(cstar_enter)
> -        /* sti could live here when we don't switch page tables below. */
> +        ALTERNATIVE nop, sti, X86_FEATURE_NO_XPTI

I do not think the complexity of of altering the position of sti
outweighs the fractional extra delay which would result from
unilaterally having the sti later.  Furthermore, if you really are
concerned about microoptimising this, you don't want a singlebyte nop here.

>          CR4_PV32_RESTORE
>          movq  8(%rsp),%rax /* Restore %rax. */
>          movq  $FLAT_KERNEL_SS,8(%rsp)
> @@ -201,6 +201,7 @@ ENTRY(cstar_enter)
>          SAVE_ALL
>  
>          GET_STACK_END(bx)
> +.Lcstar_cr3_start:
>          mov   STACK_CPUINFO_FIELD(xen_cr3)(%rbx), %rcx
>          neg   %rcx
>          jz    .Lcstar_cr3_okay
> @@ -210,6 +211,12 @@ ENTRY(cstar_enter)
>          movq  $0, STACK_CPUINFO_FIELD(xen_cr3)(%rbx)
>  .Lcstar_cr3_okay:
>          sti
> +.Lcstar_cr3_end:
> +        .pushsection .altinstructions, "a", @progbits
> +        altinstruction_entry .Lcstar_cr3_start, .Lcstar_cr3_start, \
> +                             X86_FEATURE_NO_XPTI, \
> +                             (.Lcstar_cr3_end - .Lcstar_cr3_start), 0
> +        .popsection

It occurs to me that this would be far more legible if we had an alt_nop
wrapper.  Reusing .Lcstar_cr3_start and a length of 0 isn't obvious.

> --- a/xen/arch/x86/x86_64/entry.S
> +++ b/xen/arch/x86/x86_64/entry.S
> @@ -46,7 +47,6 @@ restore_all_guest:
>          movabs $DIRECTMAP_VIRT_START, %rcx
>          mov   %rdi, %rax
>          and   %rsi, %rdi
> -        jz    .Lrag_keep_cr3

This looks like a functional change?

>          and   %r9, %rsi
>          add   %rcx, %rdi
>          add   %rcx, %rsi
> @@ -473,6 +499,7 @@ ENTRY(dom_crash_sync_extable)
>  ENTRY(common_interrupt)
>          SAVE_ALL CLAC
>  
> +.Lintr_cr3_start:
>          GET_STACK_END(14)
>          mov   STACK_CPUINFO_FIELD(xen_cr3)(%r14), %rcx
>          mov   %rcx, %r15
> @@ -492,9 +519,20 @@ ENTRY(common_interrupt)
>          CR4_PV32_RESTORE
>          movq %rsp,%rdi
>          callq do_IRQ
> +.Lintr_cr3_restore:
>          mov   %r15, STACK_CPUINFO_FIELD(xen_cr3)(%r14)
> +.Lintr_cr3_end:
>          jmp ret_from_intr
>  
> +        .pushsection .altinstructions, "a", @progbits
> +        altinstruction_entry .Lintr_cr3_restore, .Lintr_cr3_restore, \
> +                             X86_FEATURE_NO_XPTI, \
> +                             (.Lintr_cr3_end - .Lintr_cr3_restore), 0
> +        altinstruction_entry .Lintr_cr3_start, .Lintr_cr3_start, \
> +                             X86_FEATURE_NO_XPTI, \
> +                             (.Lintr_cr3_okay - .Lintr_cr3_start), 0

This is now getting very complicated to follow.  Is it just for IST
safety and liable to disappear?  If not, I think we need a different
way,as this is now saying "sporadic instructions inside this block, but
not all of them, turn into nops".

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel