[RFC] ptrace: add generic SET_SYSCALL request

[arnd@arndb.de (Arnd Bergmann)]

On Thursday 13 November 2014 16:02:49 AKASHI Takahiro wrote:
> On 11/12/2014 08:19 PM, Arnd Bergmann wrote:
> > On Wednesday 12 November 2014 11:13:52 Will Deacon wrote:
> >> On Wed, Nov 12, 2014 at 11:06:59AM +0000, AKASHI Takahiro wrote:
> >>> On 11/12/2014 08:00 PM, Will Deacon wrote:
> >>>> On Wed, Nov 12, 2014 at 10:46:01AM +0000, AKASHI Takahiro wrote:
> >>>>> On 11/07/2014 11:04 PM, Oleg Nesterov wrote:
> >>>>>> To me the fact that PTRACE_SET_SYSCALL can be undefined and syscall_set_nr()
> >>>>>> is very much arch-dependant (but most probably trivial) means that this  code
> >>>>>> should live in arch_ptrace().
> >>>>>
> >>>>> Thinking of Oleg's comment above, it doesn't make sense neither to define generic
> >>>>> NT_SYSTEM_CALL (user_regset) in uapi/linux/elf.h and implement it in ptrace_regset()
> >>>>> in kernel/ptrace.c with arch-defined syscall_(g)set_nr().
> >>>>>
> >>>>> Since we should have the same interface on arm and arm64, we'd better implement
> >>>>> ptrace(PTRACE_SET_SYSCALL) locally on arm64 for now (as I originally submitted).
> >>>>
> >>>> I think the regset approach is cleaner. We already do something similar for
> >>>> TLS. That would be implemented under arch/arm64/ with it's own NT type.
> >>>
> >>> Okey, so arm64 goes its own way
> >>> Or do you want to have a similar regset on arm, too?
> >>> (In this case, NT_ARM_SYSTEM_CALL can be shared in uapi/linux/elf.h)
> >>
> >> Just do arm64. We already have the dedicated request for arch/arm/.
> >
> > I wonder if we should define NT_ARM64_SYSTEM_CALL to the same value
> > as NT_S390_SYSTEM_CALL (0x307), or even define it as an architecture-
> > independent NT_SYSTEM_CALL number with that value, so other architectures
> > don't have to introduce new types when they also want to implement it.
> 
> I digged into gdb code (gdb/bfd/elf.c):
> https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;a=blob;f=bfd/elf.c;h=8b207ad872a3992381e93bdfa0a75ef444651613;hb=HEAD
>    elf_parse_notes()->elfcore_grok_note()->elfcore_grok_s390_system_call()
> 
> It seems to me that NT_S390_SYSTEM_CALL(=0x307) is recognized as a s390 specific
> value (without checking for machine type). So thinking of potential conflict, it might not be
> a good idea to use this value as a common number (of NT_SYSTEM_CALL).
> It's very unlikely that a "note" section for NT_(S390_)SYSTEM_CALL appears in a coredump file, though.
>
> What do you think?

(adding Ulrich and Andreas)

This code was introduced by http://sourceware-org.1504.n7.nabble.com/rfa-s390-bfd-part-Support-extended-register-sets-td50072.html

I have to admit that I don't really understand gdb internals, but from
a first look I get the impression that it will just do the right thing
if you reuse NT_S390_SYSTEM_CALL on ARM64 with the same semantics.

If not, we should indeed have a different number for it and duplicate that
code.

	Arnd