Re: Introduce fixed sys_sync_file_range2() syscall, implement on PowerPC and ARM

[David Woodhouse <dwmw2@infradead.org>]

On Mon, 2007-06-25 at 12:09 +0100, Russell King wrote:
> On Mon, Jun 25, 2007 at 04:35:35AM -0600, Matthew Wilcox wrote:
> > On Mon, Jun 25, 2007 at 02:11:45AM -0700, Andrew Morton wrote:
> > > ABI.  How would you like "it would be nice if maintainers of oddball
> > > architectures would pay attention"?
> > 
> > If new syscalls got posted to linux-arch for discussion, I assure you,
> > we'd pay attention.
> 
> Ditto.  sys_sync_file_range() wasn't, so I think David's sentiment is
> bang on.
> 
> And as David says, we've _finally_ been round the discussion loop with
> fallocate, so in theory we now know what the issues are, and we _all_
> have a good idea how to deal with argument ordering to satisfy the
> majority.  That should mean that subsequent discussion be shorter.

Well, those who were paying attention might. We should probably add
Documentation/new-syscalls.txt with such information as...

--- 

Never add any system call without considering how its prototype works
for all architectures and for 32-bit userspace on 64-bit kernels.

- Some architectures have a limit of 6 (32-bit) argument slots.
- Some architectures must align 64-bit integers into an aligned
  pair of registers. A slot may be wasted for padding.
- S390 may not have a 64-bit integer in slots 5/6.

Where you invent a data structure for communication with userspace, be
aware of the following:

- Try to ensure your data structure will be identical for 32-bit and
  64-bit builds, if possible. That way, you avoid the need to implement
  compatibility routines for 32-bit userspace on 64-bit kernels. In
  particular, avoid the 'long' data type. Try to use explicitly sized
  types such as 'uint64_t' instead.
- Most architectures align 64-bit integers to 8 bytes, but i386 doesn't.
  If you have to implement 32-bit compatibility, make sure you get this
  right. Preferably, avoid the problem by ensuring that your 64-bit 
  integers are naturally placed with 8-byte alignment even without
  padding.

---

> At least now, if they don't, provided we build every -rc kernel as it's
> released, we can detect when new syscalls are added quickly and give
> those submitters a suitable roasting at gas mark 95.

Linus should be refusing any new system call which doesn't at _least_
handle the 32/64 compatibility issues. Explictly stating that it doesn't
need compatibility wrappers would be OK, as long as it's true -- but
just saying _nothing_ about it is bad.

-- 
dwmw2