Re: [PATCH] compat: fix compat_sys_openat and friends

[Stephen Rothwell <sfr@canb.auug.org.au>]

On Wed, 01 Feb 2006 21:56:44 -0800 (PST) "David S. Miller" <davem@davemloft.net> wrote:
>
> From: Linus Torvalds <torvalds@osdl.org>
> Date: Wed, 1 Feb 2006 21:36:40 -0800 (PST)
> 
> > Wouldn't it be _much_ better to declare the argument as a "long", since 
> > some architectures (alpha, for example) may assume that 32-bit arguments 
> > have been _sign_extended, not zero-extended.
> > 
> > Then, when the "compat_sys_xxxx()" function passes the "long" down to the 
> > _real_ function (which takes an "int"), those architectures (and only 
> > those architectures) that actually have assumptions about high bits will 
> > have the compiler automatically do the right zero- or sign-extensions at 
> > that call-site.
> 
> There is the convention that for the compat system calls all the args
> will be 32-bit zero extended by the platform syscall entry code before
> the C code is invoked.  This topic used to come up a lot and finally
> we all decided that was the thing to do.
> 
> It's important (at least I think so :-) for all of this generic compat
> code to be able to have a well defined argument environment.
> 
> Anyways, I think that's how Stephen arrived at his patch.

Yes, that is it.  I have tried using "long" and "unsigned int" for those
first parameters and it produces exactly the same assembler output on
ppc64 and x86_64.  Everywhere else that we have a file descriptor argument
to a compat syscall function it is declared "unsigned int".

And for these compat functions, alpha is irrelevent of course. :-)

-- 
Cheers,
Stephen Rothwell                    sfr@canb.auug.org.au
http://www.canb.auug.org.au/~sfr/

Attachment: pgpE1GNUAfWun.pgp
Description: PGP signature