Re: sys getdents64 needs compat wrapper ?

[Arnd Bergmann <arnd@arndb.de>]

[-- Attachment #1: Type: text/plain, Size: 1217 bytes --]

On maanantai 07 kesäkuu 2004 23:13, Arun Sharma wrote:
> This updated patch takes care of the __user problems that Dave
> pointed out. Regarding __put_user vs __copy_to_user, __put_user
> is more optimal on ia64 like Andrew already noted. I realize it's
> not optimal on all platforms. So I suggest that we apply the 
> patch for now and try to build something that's optimal for all
> platforms in the future.     

> +       if (dirent) {
> +               if ((__put_user(offset, (u32 __user *)&dirent->d_off))
> +                || (__put_user(offset >> 32, ((u32 __user *)&dirent->d_off) + 1)))
> +                       goto efault;
> +       }

I just realized your code is broken on big-endian architectures, so
we should either get back to __copy_to_user or find the optimal solution
now.

How about a {__,}{get,put}_user_unaligned() that gets defined per
architecture in one of these two ways:

/* put_user doesn't care about alignment */
#define put_user_unaligned(x,ptr) put_user(x,ptr)

/* put_user only works on aligned data */
#define put_user_unaligned(x,ptr) ({ \
	__typeof__(*(ptr)) __x = (x); \
	copy_to_user(ptr, &__x, sizeof(*(ptr))) ? -EFAULT : 0; \
})

	Arnd <><

[-- Attachment #2: signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]