Re: [musl] [PATCH] arm64/sigcontext: Synchronize the type of the __reserved field with the linux kernel.

[Fangrui Song <i@maskray.me>]

On Thu, Aug 19, 2021 at 11:46 PM Szabolcs Nagy <nsz@port70.net> wrote:
>
> * Rich Felker <dalias@libc.org> [2021-08-19 09:38:41 -0400]:
> > On Thu, Aug 19, 2021 at 07:54:12AM +0200, Florian Weimer wrote:
> > > * Rich Felker:
> > >
> > > > On Thu, Aug 19, 2021 at 12:52:23AM +0200, Olivier Galibert wrote:
> > > >> clang's compiler-rt sanitizer_linux.cpp expects the __reserved field
> > > >> to be convertible to u8 *.  So let's.
> > > >> ---
> > > >>  arch/aarch64/bits/signal.h | 2 +-
> > > >>  1 file changed, 1 insertion(+), 1 deletion(-)
> > > >>
> > > >> diff --git a/arch/aarch64/bits/signal.h b/arch/aarch64/bits/signal.h
> > > >> index 5098c734..a46997e3 100644
> > > >> --- a/arch/aarch64/bits/signal.h
> > > >> +++ b/arch/aarch64/bits/signal.h
> > > >> @@ -19,7 +19,7 @@ typedef struct sigcontext {
> > > >>          unsigned long fault_address;
> > > >>          unsigned long regs[31];
> > > >>          unsigned long sp, pc, pstate;
> > > >> -        long double __reserved[256];
> > > >> +        unsigned char __reserved[4096] __attribute__((__aligned__(16)));
> > > >>  } mcontext_t;
> > > >>
> > > >>  #define FPSIMD_MAGIC 0x46508001
> > > >
> > > > The member name __reserved is not API, much less its particular type.
> > >
> > > The name is called __reserved, but it is actually part of the API.
> > > We learned this when we tried to rename it:
> > >
> > >   <https://sourceware.org/bugzilla/show_bug.cgi?id=22742>
> > >
> > > The name and its __ prefix are rather unfortunate, but we are stuck with
> > > it.
> >
> > I question the reasoning there. Just because there are users of it
> > doesn't mean it's API, *especially* if the users are things like
> > sanitizer lib that regularly poke at internals that are not interface
> > contracts. Use of a name like __reserved as API is *really* bad since
> > it could even be something like a macro for an attribute in the
> > implementation, rather than something available as a member name.
> >
> > My interpretation was that it's something like the powerpc reserved
> > space where there's a separate pointer into it, which you're supposed
> > to access it by. But that doesn't seem to be the case here, so I'm not
> > sure what the right way to access it is. Do you have a list of
> > software that's actually poking at it so we can evaluate the situation
> > better for figuring out what to do?
>
> __reserved holds sigcontext extensions so i'd expect
> code accessing the extended register state from a
> signal handler would use it (includes fp/simd regs).
>
> e.g. chromium tests seem to use it:
>
> https://source.chromium.org/search?q=uc_mcontext.__reserved
>
> but the ucontext access code in breakpad is written
> in asm so the c struct is not used in actual code.
>
> qemu seems to define its own types. and i don't
> immediately know of other tools that poke at
> sigcontext.

I added a reinterpret_cast in
https://github.com/llvm/llvm-project/commit/70986ea3d6aeacb5d10bfe3b75757611d4e7a379
(I have sufficient compiler-rt commits so they may not blame me)

compiler-rt from llvm-project 14.0.0 will work with the long double version.