Re: [PATCH 16/18] arm64: ptrace: handle ptrace_request differently for aarch32 and ilp32

[Catalin Marinas]

On Wed, Dec 07, 2016 at 09:40:13PM +0100, Arnd Bergmann wrote:
> On Wednesday, December 7, 2016 4:59:13 PM CET Catalin Marinas wrote:
> > On Tue, Dec 06, 2016 at 11:55:08AM +0530, Yury Norov wrote:
> > > On Mon, Dec 05, 2016 at 04:34:23PM +0000, Catalin Marinas wrote:
> > > > On Fri, Oct 21, 2016 at 11:33:15PM +0300, Yury Norov wrote:
> > > > > New aarch32 ptrace syscall handler is introduced to avoid run-time
> > > > > detection of the task type.
> > > > 
> > > > What's wrong with the run-time detection? If it's just to avoid a
> > > > negligible overhead, I would rather keep the code simpler by avoiding
> > > > duplicating the generic compat_sys_ptrace().
> > > 
> > > Nothing wrong. This is how Arnd asked me to do. You already asked this
> > > question: http://lkml.iu.edu/hypermail/linux/kernel/1604.3/00930.html
> > 
> > Hmm, I completely forgot about this ;). There is still an advantage to
> > doing run-time checking if we avoid touching core code (less acks to
> > gather and less code duplication).
> > 
> > Let's see what Arnd says but the initial patch looked simpler.
> 
> I don't currently have either version of the patch in my inbox
> (the archive is on a different machine), but in general I'd still
> think it's best to avoid the runtime check for aarch64-ilp32
> altogether. I'd have to look at the overall kernel source to
> see if it's worth avoiding one or two instances though, or
> if there are an overwhelming number of other checks that we
> can't avoid at all.

Just in case you haven't found them already, current version:

https://marc.info/?l=linux-arm-kernel&m=147708276818318&w=2

Original version:

https://patchwork.kernel.org/patch/7980521/

The old one looks more readable and given that ptrace is not really a
fast path, I'm not two worried about run-time checks

> Regarding ptrace, I notice that arch/tile doesn't even use
> the compat entry point for its ilp32 user space on 64-bit
> kernels, it just calls the regular 64-bit one. Would that
> help here?

I don't know whether it would work, we have incompatible siginfo_t on
AArch64/ILP32.

-- 
Catalin