[RFC PATCH 1/2] arm64: fpsimd: Fix bad si_code for undiagnosed SIGFPE

[Eric W. Biederman]

Dave Martin <Dave.Martin at arm.com> writes:

> On Mon, Jan 22, 2018 at 03:13:08PM -0600, Eric W. Biederman wrote:
>> 
>> > diff --git a/include/uapi/asm-generic/siginfo.h b/include/uapi/asm-generic/siginfo.h
>> > index e447283..77edb00 100644
>> > --- a/include/uapi/asm-generic/siginfo.h
>> > +++ b/include/uapi/asm-generic/siginfo.h
>> > @@ -193,7 +193,8 @@ typedef struct siginfo {
>> >  #define FPE_FLTRES	6	/* floating point inexact result */
>> >  #define FPE_FLTINV	7	/* floating point invalid operation */
>> >  #define FPE_FLTSUB	8	/* subscript out of range */
>> > -#define NSIGFPE		8
>> > +#define FPE_UNKNOWN	9	/* undiagnosed floating-point exception */
>> > +#define NSIGFPE		9
>> 
>> Minor nit here.
>> 
>> At least before this is final I would really appreciate if you could
>> rebase this on top of my unificiation of siginfo.h that I posted on
>> linux-arch and is in my siginfo-next branch.
>> 
>> As that already pushes NSIGFPE up to 13.
>> 
>> Which would make this patch change NSIGFPE to 14 and allocate the number
>> 14 for FPE_UNKNOWN
>
> Looking at this, I note a few things:
>
>  * For consistent naming, FPE_FLTUNK might be a better name for
>    FPE_UNKNOWN.
>
>    FPE_FLTUNK seems generic, tempting me to insert it as number 9
>    (since only the numbers up to 8 are ABI just now).

Except on ia64 and frv.  And who knows we might need it on one of those
architectures as well.

>    (The temptation to call it FPE_FLUNK is strong, but I can't argue
>    that's consistent...)

I totally understand the temptation.

>  * No distinction is drawn between generic and arch-dependent codes
>    here, so NSIGFPE will typically be too big.  The generic siginfo
>    handling code can detect random garbage in si_code this way, but
>    off-by-ones or misused arch-specific codes may slip through.
>
>    In particular, new x86-specific FPE_* codes will likely be
>    invisible to the BUILD_BUG_ON()s in arch/x86/kernel/signal_compat.c
>    unless so many are added that x86 overtakes ia64.

Long ago in a far off time, we had arch dependent system call numbers
and the like because that provided ABI compatibility with the existing
unix on the platform.

I don't see any of that with the siginfo si_codes.   In most cases
they are arch dependent extensions which is silly.  We should have
unconditionally extended the si_codes for all architectures in case
another architecture needs that si_code.

The fact we now have battling meanings for si_codes depending on the
architecture is an unfortunate mess.

So to me it looks most maintainable going forward to declare that all
si_codes should be allocated generically, from the same number space,
in the same header file.  While we live with the existing historic
mess.

>  * Should we reserve space for future generic codes (say up to 15)?
>    Downside: si_code validation is not a simple matter of checking
>    <= NSIGFPE in that case.  (Though <= is still better than no
>    check at all, and no worse than the current situation.)

I think new si_codes should be allocated where there are not conflicts
on any architecture.  Just in case they are useful on another
architecture in the future.

>  * What are NSIGFPE etc. doing in this header?  These aren't specified
>    by POSIX and I'm not sure what userspace would legitimately use them
>    for... though it may be too late to change this now.
>
>    Most instances on codeseaarch.debian.net are the kernel, copies
>    of kernel headers, and translated versions of kernel headers.
>    It's hard to be exhaustive though.
>
>
> We could have something like this:
>
> #define FPE_FLTUNK	9
> #define __NSIGFPE_GENERIC	9
> #define NSIGFPE		__NSIGFPE_GENERIC
>
> /* si_code <= 15 reserved for arch-independent codes */
>
> #if defined(__frv__)
>
> # define FPE_MDAOF	16
> # undef NSIGFPE
> # define NSIGFPE	16
>
> #elif define(__ia64__)
>
> # define __FPE_DECOVF	16
> # define __FPE_DECDIV	17
> # define __FPE_DECERR	18
> # define __FPR_INVASC	19
> # undef NSIGFPE
> # define NSIGFPE	19
>
> #endif
>
> (Avoiding a (base + offset) approach for the arch codes, since that
> would make it look like the codes can be renumbered safely without
> breaking anything).
>
> The generic vs. arch vs. NSIGFOO problem already exists for other
> signals.  We could take a similar approach for those, but OTOH it
> may just not be worth the effort.

What I have tried to do in my merger is discurage the idea that there
are any arch specific si_codes.  To set NSIGXXX to the largest value
from any of the architectures.  And to encourage new si_codes get
allocated after the current NSIGXXX.  So that they will work on all
architectures.

It is all a bit of a mess, but one unified mess seems like the best we
can do right now.

Eric