[PATCH] arm64: Fix off-by-one vdso trampoline return value

[William Mcvicker]

Hi All,

After digging into this even deeper with Nick, we found that the underlying
issue is with ld.lld not carrying over the st_type for the VDSO_compat* symbol
assignments in `arch/arm64/kernel/vdso32/vdso.lds.S`.

>From my Android Common Kernel 4.19 build:
$ llvm-readelf -s arch/arm64/kernel/vdso32/vdso.so.raw | grep thumb | sort
    26: 0000094d     0 NOTYPE  LOCAL  DEFAULT    11 VDSO_compat_sigreturn_thumb
    28: 00000955     0 NOTYPE  LOCAL  DEFAULT    11 VDSO_compat_rt_sigreturn_thumb
    37: 0000094d     4 FUNC    GLOBAL DEFAULT    11 __kernel_sigreturn_thumb
    38: 00000955     4 FUNC    GLOBAL DEFAULT    11 __kernel_rt_sigreturn_thumb
     8: 0000094d     4 FUNC    GLOBAL DEFAULT    11 __kernel_sigreturn_thumb@@LINUX_2.6
     9: 00000955     4 FUNC    GLOBAL DEFAULT    11 __kernel_rt_sigreturn_thumb@@LINUX_2.6

Fortunately, this has been fixed by llvm here: https://reviews.llvm.org/D86263.
So for kernels that use ld.lld and the VDSO compat sigreturn trampoline, they
need to make sure to upgrade their toolchain.

I hope this thread helps anyone running into this issue in the future.

Thanks,
Will

On 11/12/2020, William Mcvicker wrote:
> Hi Nick,
> 
> Regarding llvm-nm, this extra thumb +1 is noticed after porting
> https://lore.kernel.org/linux-arm-kernel/20201013033947.2257501-1-natechancellor@gmail.com/
> to the Android Common Kernel android-4.19-stable. I'm not sure why using ld.lld
> causes this difference, but this proposed patch ensures that we don't rely on
> the nm tool used.
> 
> Will D.,
> Regarding applying this to some stable kernels vs backporting 2d071968a405
> ("arm64: compat: Remove 32-bit sigreturn code from the vDSO"), I am hesitant to
> backport commit 2d071968a405 due it's dependencies. For 4.19 at least, I would
> also need to backport these:
> 
> 8e411be6aad13 will at kernel.org  arm64: compat: Always use sigpage for sigreturn trampoline
> a39060b009ca0 will at kernel.org  arm64: compat: Allow 32-bit vdso and sigpage to co-exist
> 1d09094aa6205 mark.rutland at arm.com  arm64: vdso: use consistent 'map' nomenclature
> d3418f3839b66 mark.rutland at arm.com  arm64: vdso: use consistent 'abi' nomenclature
> 3ee16ff3437ca mark.rutland at arm.com  arm64: vdso: simplify arch_vdso_type ifdeffery
> 74fc72e77dc5c mark.rutland at arm.com  arm64: vdso: remove aarch32_vdso_pages[]
> 
> I have done this in my local tree and verified it fixes the SIGBUS error I'm
> seeing; however, it seems a lot cleaner and safer to just patch the VDSO_SYMBOL
> macro.  Please let me know what route you prefer. I'm happy to backport all of
> these if that's the recommended approach.
> 
> Thanks,
> Will
> 
> On 11/12/2020, Will Deacon wrote:
> > On Thu, Nov 12, 2020 at 12:14:22AM +0000, Will McVicker wrote:
> > > Depending on your host nm version, the generated header
> > > `include/generated/vdso32-offsets.h` may have the bottom bit set for the
> > > thumb vdso offset addresses (as observed when using llvm-nm). This
> > > results in an additional +1 for thumb vdso trampoline return values
> > > since compat_setup_return() already includes `vdso_trampoline + thumb`.
> > > As a result, I see a SIGBUS error when running the LTP test
> > > syscalls.rt_sigaction01. To fix this, let's clear the bottom bit of the
> > > vdso_offset in the VDSO_SYMBOL macro.
> > > 
> > > Test: LTP test syscalls.rt_sigaction01
> > > Fixes: f01703b3d2e6 ("arm64: compat: Get sigreturn trampolines from vDSO")
> > > Signed-off-by: Will McVicker <willmcvicker at google.com>
> > > ---
> > >  arch/arm64/include/asm/vdso.h | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/arch/arm64/include/asm/vdso.h b/arch/arm64/include/asm/vdso.h
> > > index f99dcb94b438..a7384379e8e1 100644
> > > --- a/arch/arm64/include/asm/vdso.h
> > > +++ b/arch/arm64/include/asm/vdso.h
> > > @@ -23,7 +23,7 @@
> > >  
> > >  #define VDSO_SYMBOL(base, name)						   \
> > >  ({									   \
> > > -	(void *)(vdso_offset_##name - VDSO_LBASE + (unsigned long)(base)); \
> > > +	(void *)((vdso_offset_##name & ~1UL) - VDSO_LBASE + (unsigned long)(base)); \
> > 
> > I don't think we need this in mainline, because the sigreturn trampoline
> > is just a bunch of .byte directives and I removed the sigreturn code from
> > the compat vdso in 2d071968a405 ("arm64: compat: Remove 32-bit sigreturn code
> > from the vDSO").
> > 
> > Might be needed in some stable kernels though (or we just backport the
> > patch I mentioned above)
> > 
> > Will