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

[Will Deacon]

Hi Anisse, Dave,

On Fri, Jun 14, 2019 at 01:22:22PM +0200, Anisse Astier wrote:
> 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:
> > > -#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.

x86 appears to have an asm/prctl.h implementation, but it's not included
by anybody so I guess that doesn't really help us here.

> 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.

Certainly not a big issue, just that the harder we make this to change
the better.

> 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 ?

Duplication is grotty, but it does the job so I'm ok with it. I don't have
any better ideas.

Thanks,

Will