Re: [PATCH v3] arm64: sysreg: make mrs_s and msr_s macros work with Clang and LTO

From: Arnd Bergmann
Date: Tue Apr 23 2019 - 14:39:47 EST


On Tue, Apr 23, 2019 at 8:25 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Tue, Apr 23, 2019 at 11:15 AM Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
> > On Tue, Apr 23, 2019 at 7:12 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > >
> > > On Mon, Apr 22, 2019 at 10:44:10AM -0700, Nick Desaulniers wrote:
> > > > On Tue, Apr 16, 2019 at 9:03 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Tue, Apr 16, 2019 at 12:08 PM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > > > > >
> > > > > > On Mon, Apr 15, 2019 at 10:22:27AM -0700, Nick Desaulniers wrote:
> > > > > > > On Mon, Apr 15, 2019 at 10:06 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> > > > > > > > It would be nice if we could simply rely on a more recent binutils these
> > > > > > > > days, which supports the generic S<op0>_<op1>_<cn>_<Cm>_<op2> sysreg
> > > > > > > > definition. That would mean we could get rid of the whole msr_s/mrs_s
> > > > > > > > hack by turning that into a CPP macro which built that name.
> > > >
> > > > Mark, can you give me a test case for this? I'd like to check if
> > > > clang's integrated assembler supports this or not, so we can file an
> > > > issue and fix it if not.
> > >
> > > The below is a smoke test. The entire op0:op1:cn:cm:op2 space is 14
> > > bits, so it should be feasible to generate and run an exhautive test for
> > > all encodings.
> > >
> > > I've included SYS and SYSL here too, since I strongly suspect we'll need
> > > those going forward.
> > >
> > > Thanks
> > > Mark.
> > >
> > > ---->8----
> > > /*
> > > * ${CC} -c test.c
> > > *
> > > * OLD_* encodings already exist, NEW_* encodings haven't yet be
> > > * allocated (per ARM DDI 0487D.a), but should assemble. All encodings
> > > * chosen arbitrarily.
> > > */
> > >
> > > // OSDTRRX_EL1
> > > #define OLD_DBG_REG "s2_0_c0_c0_2"
> > > #define NEW_DBG_REG "s2_7_c0_c15_0"
> > >
> > > // MIDR_EL1
> > > #define OLD_SYS_REG "s3_0_c0_c0_0"
> > > #define NEW_SYS_REG "s3_6_c1_c0_7"
> > >
> > > #define REG(r) \
> > > asm volatile("msr " r ", xzr\n"); \
> > > asm volatile("mrs xzr, " r "\n")
> > >
> > > // DC IVAC, XZR
> > > #define OLD_SYS "#0, c7, c6, #1"
> > > #define NEW_SYS "#7, c15, c15, #7"
> > >
> > > #define NEW_SYSL_MIN "#0, c0, c0, #0"
> > > #define NEW_SYSL_MAX "#7, c15, c15, #7"
> > >
> > > #define SYS(s) asm volatile("sys " s ", xzr\n")
> > > #define SYSL(s) asm volatile("sysl xzr, " s "\n")
> > >
> > > void test(void)
> > > {
> > > REG(OLD_DBG_REG);
> > > REG(NEW_DBG_REG);
> > >
> > > REG(OLD_SYS_REG);
> > > REG(NEW_SYS_REG);
> > >
> > > SYS(OLD_SYS);
> > > SYS(NEW_SYS);
> > >
> > > SYSL(NEW_SYSL_MIN);
> > > SYSL(NEW_SYSL_MAX);
> > > }
> >
> > Hmmm...yes looks like Clang's integrated assembler is missing support.
> > https://godbolt.org/z/wDBpz4

It looks like you forgot to pass --target=aarch64-linux-gnu to clang here,
which produces a build failure. This is the output with the added flag:

https://godbolt.org/z/HYoEH1

> > https://bugs.llvm.org/show_bug.cgi?id=41573
> > https://github.com/ClangBuiltLinux/linux/issues/453
> > We'll follow up on getting that fixed.
> >
> > Kees, are you able to respin the patch with the above requested
> > changes from Mark to the macros? Let me know what I can do to help.
>
> Doesn't this mean arm64 must raise their minimum binutils to 2.25? Is
> that acceptable? I hadn't seen that question answered yet. (Maybe I
> missed it?)

Generally I'd say that's a bit too new for a minimum version. On the other
hand, most distros I could find for arm64 have at least 2.25 or later,
the oldest
one I found being Debian Jessie. Ubuntu Trusty (14.04) was using 2.24 but
has its end of support date this month, after five years.

However, there are probably still cross-compile toolchains around that use
older aarch64-binutils, so it probably hurts someone.

Arnd