Re: [PATCH 0/3] arm64/sve: UAPI: Disentangle ptrace.h from sigcontext.h

[Dave P Martin <Dave.Martin@arm.com>]

On Fri, Dec 14, 2018 at 07:00:07PM +0000, Szabolcs Nagy wrote:
> On 14/12/2018 18:25, Dave Martin wrote:
> > On Fri, Dec 14, 2018 at 06:13:33PM +0000, Szabolcs Nagy wrote:
> >> On 11/12/2018 19:26, Dave Martin wrote:
> >>> This patch refactors the UAPI header definitions for the Arm SVE
> >>> extension to avoid multiple-definition problems when userspace mixes its
> >>> own sigcontext.h definitions with the kernel's ptrace.h (which is
> >>> apparently routine).
> >>>
> >>> A common backend header is created to hold common definitions, suitably
> >>> namespaced, and with an appropriate header guard.
> >>>
> >>> See the commit message in patch 3 for further explanation of why this
> >>> is needed.
> >>>
> >>> Because of the non-trivial header guard in the new sve_context.h, patch
> >>> 1 adds support to headers_install.sh to munge #if defined(_UAPI_FOO) in
> >>> a similar way to the current handling of #ifndef _UAPI_FOO.
> >>>
> >>
> >> thanks for doing this.
> >>
> >> the patches fix the gdb build issue on musl libc with an
> >> additional gdb patch:
> >> https://sourceware.org/ml/gdb-patches/2018-12/msg00152.html
> >> (in userspace i'd expect users relying on signal.h providing
> >> whatever is in asm/sigcontext.h.)
> >>
> >> i think sve_context.h could be made to work with direct include,
> >> even if that's not useful because there is no public api there.
> >> (and then you dont need the first patch)
> > 
> > My general view is that if you want the sigframe types userspace should
> > usually include <ucontext.h> and refer to mcontext_t.
> > 
> 
> ucontext.h does not expose the asm/sigcontext.h types in glibc,
> but it is compatible with the inclusion of asm/sigcontext.h
> (or signal.h).
> 
> in musl ucontext.h includes signal.h and signal.h provides
> the asm/sigcontext.h api with abi compatible definitions.
> 
> > Because the prototype for sa_sigaction() specifies a void * for the
> > ucontext argument, I've generally assumed that <signal.h> is not
> > sufficient to get ucontext_t (or mcontext_t) (but maybe I'm too paranoid
> > there).
> 
> http://pubs.opengroup.org/onlinepubs/9699919799/basedefs/signal.h.html
> 
> "The <signal.h> header shall define the ucontext_t type as a structure
>  that shall include at least the following members:
> ...
>  mcontext_t  uc_mcontext A machine-specific representation of the saved
>              context."
> 
> so signal.h must define ucontext_t but mcontext_t can be opaque.
> (it is opaque with posix conform feature tests to avoid
> namespace pollution, but with _GNU_SOURCE defined all
> asm/sigcontext.h apis are there and mcontext_t matches
> struct sigcontext)

I see.  Sounds reasonable.

> > 
> > Non-POSIX-flavoured software might include <asm/sigcontext.h> directly.
> > In glibc/musl libc will that conflict with <signal.h>, or can the two
> > coexist?
> 
> if you compile e.g in standard conform mode then
> i think signal.h and asm/sigcontext.h are compatible.

So long as we don't break any existing usage (?) I guess this is fine.

Cheers
---Dave