On Tue, Feb 02, 2016 at 08:54:34PM +0100, Heiko Carstens wrote:

Hi Yury,

On Tue, Feb 02, 2016 at 05:08:26PM +0100, Heiko Carstens wrote:

quoted

See e.g. 485d52768685 ("sys_personality: change sys_personality() to accept
"unsigned int" instead of u_long") would have been a candidate which could
silently break architectures which need compat wrappers.

Ok, this example is of course wrong. But now I can claim that also somebody
who should know better makes these mistakes.. :)

Yep, this is a bad example. :) Moreover, this patch is coming from year 2010,
and it shows how stable the syscall ABI is.

quoted

quoted

I don't know much about s390 specifics. Maybe because of that I do not
understand completely your worries. I'm OK with both 1st and 2nd
version, but I'd choose 2nd one because it allows inlines, and we
don't need the compat_wrapper.c.

It would be only nicer if we can guarentee correctness all the time. That
being said I'm about to revert my own commit :)

So if you want to go without compat_wrapper.c then we should have a
solution which will do the right thing all the time without that a system
call author has to know about the sign and zero extension issue some
architectures face. It _will_ go wrong.

So I think I can summarize my point to: if you can enforce correctness, why
shouldn't you do it if the performance impact is only a single instruction.

For aarch64 it's 5 instructions. But what's more important (if ever),
another wrapper takes another i-cache line...
        <compat_SyS_ftruncate>:
        stp     x29, x30, [sp,#-16]!
        mov     x29, sp
        bl      d40 <do_sys_ftruncate.constprop.3>
        ldp     x29, x30, [sp],#16
        ret

However I'll try to write an addon patch to your patch series. Maybe we can
still get rid of compat_wrapper.c in a way which makes both of us happy.
Also.. the idea with the alias names for compat wrappers does seem to have
the disadvantage that it will pollute /proc/kallsyms for example.

Anyway, I'm not sure if I will be able to come up with something this week
though.

Great, I'm looking forward...

My point is. Syscall ABI is so stable and so important that too much
people are involved in development and testing of it. So automatic
checks are almost useless, as all bugs will be found during
development and review. I think, community is able to pay enough
attention to review a couple of such patches per decade.
But I'm OK with some automatic checker, if it will be :
 - not too complicated, as complex code makes bugs by itself;
 - not too expensive;
 - not too ugly (my one is definitely ugly).