[PATCH] KVM: arm64: Disable LTO in hyp

[Marc Zyngier]

On 2021-03-05 16:55, Sami Tolvanen wrote:
> On Fri, Mar 5, 2021 at 6:22 AM Ard Biesheuvel <ardb at kernel.org> wrote:
>> 
>> On Fri, 5 Mar 2021 at 12:38, Marc Zyngier <maz at kernel.org> wrote:
>> >
>> > On Fri, 05 Mar 2021 02:38:17 +0000,
>> > Sami Tolvanen <samitolvanen at google.com> wrote:
>> > >
>> > > On Thu, Mar 4, 2021 at 2:34 PM Sami Tolvanen <samitolvanen at google.com> wrote:
>> > > >
>> > > > On Thu, Mar 4, 2021 at 2:17 PM Marc Zyngier <maz at kernel.org> wrote:
>> > > > >
>> > > > > On Thu, 04 Mar 2021 21:25:41 +0000,
>> > > > > Sami Tolvanen <samitolvanen at google.com> wrote:
>> >
>> > [...]
>> >
>> > > > > > I assume hyp_panic() ends up being placed too far from __guest_enter()
>> > > > > > when the kernel is large enough. Possibly something to do with LLVM
>> > > > > > always splitting functions into separate sections with LTO. I'm not
>> > > > > > sure why the linker cannot shuffle things around to make everyone
>> > > > > > happy in this case, but I confirmed that this patch also fixes the
>> > > > > > build issue for me:
>> > > > > >
>> > > > > > diff --git a/arch/arm64/kvm/hyp/vhe/switch.c b/arch/arm64/kvm/hyp/vhe/switch.c
>> > > > > > index af8e940d0f03..128197b7c794 100644
>> > > > > > --- a/arch/arm64/kvm/hyp/vhe/switch.c
>> > > > > > +++ b/arch/arm64/kvm/hyp/vhe/switch.c
>> > > > > > @@ -214,7 +214,7 @@ static void __hyp_call_panic(u64 spsr, u64 elr, u64 par)
>> > > > > >  }
>> > > > > >  NOKPROBE_SYMBOL(__hyp_call_panic);
>> > > > > >
>> > > > > > -void __noreturn hyp_panic(void)
>> > > > > > +void __noreturn hyp_panic(void) __section(".text")
>> > > > > >  {
>> > > > > >         u64 spsr = read_sysreg_el2(SYS_SPSR);
>> > > > > >         u64 elr = read_sysreg_el2(SYS_ELR);
>> > > > > >
>> > > > >
>> > > > > We're getting into black-magic territory here. Why wouldn't hyp_panic
>> > > > > be in the .text section already?
>> > > >
>> > > > It's not quite black magic. LLVM essentially flips on
>> > > > -ffunction-sections with LTO and therefore, hyp_panic() will be in
>> > > > .text.hyp_panic in vmlinux.o, while __guest_enter() will be in .text.
>> > > > Everything ends up in .text when we link vmlinux, of course.
>> > > >
>> > > > $ readelf --sections vmlinux.o | grep hyp_panic
>> > > >   [3936] .text.hyp_panic   PROGBITS         0000000000000000  004b56e4
>> > >
>> > > Note that disabling LTO here has essentially the same effect as using
>> > > __section(".text"). It stops the compiler from splitting these
>> > > functions into .text.* sections and makes it less likely that
>> > > hyp_panic() ends up too far away from __guest_enter().
>> > >
>> > > If neither of these workarounds sound appealing, I suppose we could
>> > > alternatively change hyp/entry.S to use adr_l for hyp_panic. Thoughts?
>> >
>> > That would be an actual fix instead of a workaround, as it would
>> > remove existing assumptions about the relative locations of the two
>> > objects. I guess you need to fix both instances with something such
>> > as:
>> >
>> > diff --git a/arch/arm64/kvm/hyp/entry.S b/arch/arm64/kvm/hyp/entry.S
>> > index b0afad7a99c6..a43e1f7ee354 100644
>> > --- a/arch/arm64/kvm/hyp/entry.S
>> > +++ b/arch/arm64/kvm/hyp/entry.S
>> > @@ -85,8 +85,10 @@ SYM_INNER_LABEL(__guest_exit_panic, SYM_L_GLOBAL)
>> >
>> >         // If the hyp context is loaded, go straight to hyp_panic
>> >         get_loaded_vcpu x0, x1
>> > -       cbz     x0, hyp_panic
>> > -
>> > +       cbnz    x0, 1f
>> > +       adr_l   x0, hyp_panic
>> > +       br      x0
>> > +1:
>> 
>> Agree with replacing the conditional branches that refer to external
>> symbols: the compiler never emits those, for the reason we are seeing
>> here, i.e., the range is simply insufficient.
>> 
>> But let's just use 'b hyp_panic' instead, no?
> 
> Alright, this seems to work for me:
> 
> diff --git a/arch/arm64/kvm/hyp/entry.S b/arch/arm64/kvm/hyp/entry.S
> index b0afad7a99c6..c62265951467 100644
> --- a/arch/arm64/kvm/hyp/entry.S
> +++ b/arch/arm64/kvm/hyp/entry.S
> @@ -85,8 +85,10 @@ SYM_INNER_LABEL(__guest_exit_panic, SYM_L_GLOBAL)
> 
>         // If the hyp context is loaded, go straight to hyp_panic
>         get_loaded_vcpu x0, x1
> -       cbz     x0, hyp_panic
> +       cbnz    x0, 1f
> +       b       hyp_panic
> 
> +1:
>         // The hyp context is saved so make sure it is restored to 
> allow
>         // hyp_panic to run at hyp and, subsequently, panic to run in 
> the host.
>         // This makes use of __guest_exit to avoid duplication but sets 
> the
> @@ -94,7 +96,7 @@ SYM_INNER_LABEL(__guest_exit_panic, SYM_L_GLOBAL)
>         // current state is saved to the guest context but it will only 
> be
>         // accurate if the guest had been completely restored.
>         adr_this_cpu x0, kvm_hyp_ctxt, x1
> -       adr     x1, hyp_panic
> +       adr_l   x1, hyp_panic
>         str     x1, [x0, #CPU_XREG_OFFSET(30)]
> 
>         get_vcpu_ptr    x1, x0
> 
> But when I say work, I mean this fixes the allmodconfig build with
> LTO, and my kernel boots at EL2. I don't actually have a way to
> properly test KVM on arm64. If nobody sees obvious issues here, I can
> send a proper patch a bit later.

Please do, it is "obviously correct"! ;-)

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...