Re: [UPDATED PATCH] fix memory corruption from misinterpreted bad_inode_ops return values

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

On Thu, Jan 04, 2007 at 01:30:47PM -0800, Linus Torvalds wrote:
 
> I'll happily cast away arguments that aren't used, but I'm not sure that 
> we ever should cast different return values (not "int" vs "long", but also 
> not "loff_t" etc). 
> 
> On 32-bit architectures, 64-bit entities may be returned totally different 
> ways (ie things like "caller allocates space for them and passes in a 
> magic pointer to the return value as the first _real_ argument").
> 
> So with my previous email, I was definitely _not_ trying to say that 
> casting function pointers is ok. In practice it is ok when the _arguments_ 
> differ, but not necessarily when the _return-type_ differs.
> 
> I was cc'd into the discussion late, so I didn't realize that we 
> apparently already have a situation where changing the return value to 
> "long" might make a difference. If so, I agree that we shouldn't do this 
> at all (although Andrew's change to "long" seems perfectly fine as a "make 
> old cases continue to work" patch if it actually matters).

We do.
        loff_t (*llseek) (struct file *, loff_t, int);
...
        int (*readdir) (struct file *, void *, filldir_t);

static const struct file_operations bad_file_ops =
{
        .llseek         = EIO_ERROR,
...
        .readdir        = EIO_ERROR,


Moreover, we have int, loff_t, ssize_t and long, plus the unsigned variants.
At least 3 versions, unless you want to mess with ifdefs to reduce them to
two.