Re: [PATCH] arm64: vdso32: force vdso32 to be compiled as -marm

[Peter Smith]

> From: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
> Sent: 27 May 2020 21:31
> To: Robin Murphy
> Cc: Catalin Marinas; Will Deacon; Naohiro Aota; Stephen Boyd; Masahiro Yamada; LKML; Manoj Gupta; Luis Lozano; Nathan Chancellor; Vincenzo Frascino; Linux ARM; Kristof Beyls; Victor Campos; david.spickett@xxxxxxxxxx; Arnd Bergmann; Peter Smith
> Subject: Re: [PATCH] arm64: vdso32: force vdso32 to be compiled as -marm
> 
> On Wed, May 27, 2020 at 1:14 PM Nick Desaulniers
> <ndesaulniers@xxxxxxxxxx> wrote:
> >
> > On Wed, May 27, 2020 at 12:28 PM Robin Murphy <robin.murphy@xxxxxxx> wrote:
> > >
> > > On 2020-05-27 18:55, Nick Desaulniers wrote:
> > > > On Wed, May 27, 2020 at 6:45 AM Robin Murphy <robin.murphy@xxxxxxx> wrote:
> > > >>
> > > >> On 2020-05-26 18:31, Nick Desaulniers wrote:
> > > >>> Custom toolchains that modify the default target to -mthumb cannot
> > > >>> compile the arm64 compat vdso32, as
> > > >>> arch/arm64/include/asm/vdso/compat_gettimeofday.h
> > > >>> contains assembly that's invalid in -mthumb.  Force the use of -marm,
> > > >>> always.
> > > >>
> > > >> FWIW, this seems suspicious - the only assembly instructions I see there
> > > >> are SWI(SVC), MRRC, and a MOV, all of which exist in Thumb for the
> > > >> -march=armv7a baseline that we set.
> > > >>
> > > >> On a hunch, I've just bodged "VDSO_CFLAGS += -mthumb" into my tree and
> > > >> built a Thumb VDSO quite happily with Ubuntu 19.04's
> > > >> gcc-arm-linux-gnueabihf. What was the actual failure you saw?
> > > >
> > > >  From the link in the commit message: `write to reserved register 'R7'`
> > > > https://godbolt.org/z/zwr7iZ
> > > > IIUC r7 is reserved for the frame pointer in THUMB?
> > >
> > > It can be, if you choose to build with frame pointers and the common
> > > frame pointer ABI for Thumb code that uses r7. However it can also be
> > > for other things like the syscall number in the Arm syscall ABI too.
> >
> > Ah, right, with -fomit-frame-pointer, this error also goes away.  Not
> > sure if we prefer either:
> > - build the compat vdso as -marm always or
> > - disable frame pointers for the vdso (does this have unwinding implications?)
> > - other?
> >
> > > I
> > > take it Clang has decided that writing syscall wrappers with minimal
> > > inline asm is not a thing people deserve to do without arbitrary other
> > > restrictions?
> >
> > Was the intent not obvious? We would have gotten away with it, too, if
> > wasn't for you meddling kids and your stupid dog! /s
> > https://www.youtube.com/watch?v=hXUqwuzcGeU
> > Anyways, this seems to explain more the intentions:
> > https://reviews.llvm.org/D76848#1945810
> > + Victor, Kristof (ARM)
> 
> And maybe some other useful data points regarding warning on use of r7
> and frame pointers.
> https://github.com/ClangBuiltLinux/linux/issues/701#issuecomment-591325758
> https://bugs.llvm.org/show_bug.cgi?id=45826
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94986
> 
> + Peter (ARM)
> + David, Arnd (Linaro)
> --
> Thanks,
> ~Nick Desaulniers

Hello Nick,

The AAPCS has only recently (28th January 2020) been updated with a
specification of the frame pointer and frame chain.

My understanding is that neither gcc nor clang implement this part yet.
Historically the frame pointer in Thumb has not been defined in the
AAPCS with implementations falling back to historic definitions of
fp = r7 in Thumb and fp = r11 in Arm. The use of different frame
pointer registers in Arm and Thumb state makes it difficult to
construct a frame chain with interworking. My understanding of the
AAPCS is that it has been changed to make the frame pointer r11 on
both Arm and Thumb to make unwinding possible, at the expense of r11
being harder to access from 16-bit Thumb instructions. There are 4
levels of conformance with respect to construction of frame records
and frame chain as it is likely some platforms will want to persist
with r7.

An implementation of the latest version of the AAPCS would permit
a frame pointer and use of r7 as a reserved register. Although as
you'll need to support older compilers this may not be an option.
I suggest using Arm if you need a frame pointer, and disable the
frame pointer if you want/need to use Thumb. My understanding is that
runtime unwinding using the frame pointer in Thumb is already difficult
due to Arm and Thumb functions using different registers for the frame
pointer.

Hope this helps

Peter