Re: [PATCH 5/9] fs: remove various compat readv/writev helpers

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Sep 23, 2020 at 05:38:31PM +0100, Al Viro wrote:
> On Wed, Sep 23, 2020 at 03:59:01PM +0100, Al Viro wrote:
> 
> > > That's a very good question.  But it does not just compile but actually
> > > works.  Probably because all the syscall wrappers mean that we don't
> > > actually generate the normal names.  I just tried this:
> > > 
> > > --- a/include/linux/syscalls.h
> > > +++ b/include/linux/syscalls.h
> > > @@ -468,7 +468,7 @@ asmlinkage long sys_lseek(unsigned int fd, off_t offset,
> > >  asmlinkage long sys_read(unsigned int fd, char __user *buf, size_t count);
> > >  asmlinkage long sys_write(unsigned int fd, const char __user *buf,
> > >                             size_t count);
> > > -asmlinkage long sys_readv(unsigned long fd,
> > > +asmlinkage long sys_readv(void *fd,
> > > 
> > > for fun, and the compiler doesn't care either..
> > 
> > Try to build it for sparc or ppc...
> 
> FWIW, declarations in syscalls.h used to serve 4 purposes:
> 	1) syscall table initializers needed symbols declared
> 	2) direct calls needed the same
> 	3) catching mismatches between the declarations and definitions
> 	4) centralized list of all syscalls
> 
> (2) has been (thankfully) reduced for some time; in any case, ksys_... is
> used for the remaining ones.
> 
> (1) and (3) are served by syscalls.h in architectures other than x86, arm64
> and s390.  On those 3 (1) is done otherwise (near the syscall table initializer)
> and (3) is not done at all.
> 
> I wonder if we should do something like
> 
> SYSCALL_DECLARE3(readv, unsigned long, fd, const struct iovec __user *, vec,
> 		 unsigned long, vlen);
> in syscalls.h instead, and not under that ifdef.
> 
> Let it expand to declaration of sys_...() in generic case and, on x86, into
> __do_sys_...() and __ia32_sys_...()/__x64_sys_...(), with types matching
> what SYSCALL_DEFINE ends up using.
> 
> Similar macro would cover compat_sys_...() declarations.  That would
> restore mismatch checking for x86 and friends.  AFAICS, the cost wouldn't
> be terribly high - cpp would have more to chew through in syscalls.h,
> but it shouldn't be all that costly.  Famous last words, of course...
> 
> Does anybody see fundamental problems with that?

Just to make it clear - I do not propose to fold that into this series;
there we just need to keep those declarations in sync with fs/read_write.c