Re: [PATCH] arm64/sve: <uapi/asm/ptrace.h> should not depend on <uapi/linux/prctl.h>

[Anisse Astier]

Hi Dave,

Thanks for taking the time to review this patch,

On Thu, Jun 13, 2019 at 06:14:44PM +0100, Dave Martin wrote:
> On Thu, Jun 13, 2019 at 06:38:01PM +0200, Anisse Astier wrote:
> > Otherwise this will create userspace build issues for any program
> > (strace, qemu) that includes both <sys/prctl.h> (with musl libc) and
> > <linux/ptrace.h> (which then includes <asm/ptrace.h>), like this:
> > 
> > 	error: redefinition of 'struct prctl_mm_map'
> > 	 struct prctl_mm_map {
> > 
> > See https://github.com/foundriesio/meta-lmp/commit/6d4a106e191b5d79c41b9ac78fd321316d3013c0
> > for a public example of people working around this issue.
> > 
> > This fixes an UAPI regression introduced in commit 43d4da2c45b2
> > ("arm64/sve: ptrace and ELF coredump support").
> > 
> > Cc: stable@xxxxxxxxxxxxxxx
> 
> Consider adding a Fixes: tag.

Will do in v2.

> 
> > Signed-off-by: Anisse Astier <aastier@xxxxxxxxxx>
> > ---
> >  arch/arm64/include/uapi/asm/ptrace.h | 8 +++-----
> >  1 file changed, 3 insertions(+), 5 deletions(-)
> > 
> > diff --git a/arch/arm64/include/uapi/asm/ptrace.h b/arch/arm64/include/uapi/asm/ptrace.h
> > index d78623acb649..03b6d6f029fc 100644
> > --- a/arch/arm64/include/uapi/asm/ptrace.h
> > +++ b/arch/arm64/include/uapi/asm/ptrace.h
> > @@ -65,8 +65,6 @@
> >  
> >  #ifndef __ASSEMBLY__
> >  
> > -#include <linux/prctl.h>
> > -
> >  /*
> >   * User structures for general purpose, floating point and debug registers.
> >   */
> > @@ -113,10 +111,10 @@ struct user_sve_header {
> >  
> >  /*
> >   * Common SVE_PT_* flags:
> > - * These must be kept in sync with prctl interface in <linux/ptrace.h>
> > + * These must be kept in sync with prctl interface in <linux/prctl.h>
> 
> Ack
> 
> >   */
> > -#define SVE_PT_VL_INHERIT		(PR_SVE_VL_INHERIT >> 16)
> > -#define SVE_PT_VL_ONEXEC		(PR_SVE_SET_VL_ONEXEC >> 16)
> > +#define SVE_PT_VL_INHERIT		(1 << 1) /* PR_SVE_VL_INHERIT */
> > +#define SVE_PT_VL_ONEXEC		(1 << 2) /* PR_SVE_SET_VL_ONEXEC */
> 
> Makes sense, but...
> 
> Since sve_context.h was already introduced to solve a closely related
> problem, I wonder whether we can provide shadow definitions there,
> similarly to way the arm64/include/uapi/asm/ptrace.h definitions are
> derived.  Although it's a slight abuse of that header, I think that
> would be my preferred approach.


Yes I saw this, and I considered doing something similar. But, those
defines are in uapi/linux/prctl.h, which does not include any asm/*.h
header. This would have then required adding a full infrastructure for
asm/prctl.h (that could then include sve_context.h for example), which
does not exist yet, instead of copying these two values.

Since this is part of the kernel-userspace ABI, I don't see this values
changing anytime soon, which is why I thought copying them shouldn't be
a big issue.

A simple solution would be to to include sve_context.h or a third
header, maybe linux/prctl_arm64_sve.h (with only these two/five
defines), in linux/prctl.h, and reuse it in uapi/asm/ptrace.h; but this
would break the self-contained nature of linux/prctl.h.

> 
> Otherwise, at least make the required relationship between ptrace.h and
> prctl.h constants a bit more obvious, say,
> 
> 	#define SVE_PT_VL_INHERIT ((1 << 17) /* PR_SVE_SET_VL_INHERIT */ >> 16)

This one is much simpler and closer to what I had in mind with this
patch.

Will, what do you think of this second approach Dave proposed ?

Regards,

Anisse