On Sun, Mar 18, 2018 at 11:06:42AM -0700, Linus Torvalds wrote:

and then we can do

  COMPAT_SYSCALL_DEFINE5(readahead, int, fd,
COMPAT_ARG_64BIT_ODD(off), compat_size_t, count)
  {
      return do_readahead(fd, off_lo + ((u64)off_hi << 64), count);
  }

which at least looks reasonably legible, and has *zero* ifdef's anywhere.

It's a bit more complicated, but...

I do *not* want to see those disgusting __ARCH_WANT_LE_COMPAT_SYS
things and crazy #ifdef's in code.

Absolutely.  Those piles of ifdefs are unreadable garbage.

So either let the architectures do their own trivial wrappers
entirely, or do something clean like the above. Do *not* do
#ifdef'fery at the system call declaration time.

Also note that the "ODD" arguments may not be the ones that need
padding. I could easily see a system call argument numbering scheme
like

   r0 - system call number
   r1 - first argument
   r2 - second argument
   ...

and then it's the *EVEN* 64-bit arguments that would need the padding
(because they are actually odd in the register numbers). The above
COMPAT_ARG_64BIT[_ODD]() model allows for that too.

Of course, if some architecture then has some other arbitrary rules (I
could see register pairing rules that aren't the usual "even register"
ones), then such an architecture would really have to have its own
wrapper, but the above at least would handle the simple cases, and
doesn't look disgusting to use.

I'd done some digging in that area, will find the notes and post.
Basically, we can even avoid the odd/even annotations and have
COMPAT_SYSCALL_DEFINE... sort it out.  It's a bit more hairy than
I would like at this stage in the cycle, though.  I'll see if it can
be done without too much PITA.

However, there still are genuinely speci^Wfucked in head cases - see
e.g. this sad story:
commit ab8a261ba5e2dd9206da640de5870cc31d568a7c
Author: Helge Deller [off-list ref]
Date:   Thu Jul 10 18:07:17 2014 +0200

    parisc: fix fanotify_mark() syscall on 32bit compat kernel

Those certainly ought to stay in arch/*