On Tue, Dec 28, 2021 at 3:39 PM <guoren@xxxxxxxxxx> wrote:
>
> From: Guo Ren <guoren@xxxxxxxxxxxxxxxxx>
>
> Remove duplicate F_GETLK64,F_SETLK64,F_SETLKW64 definitions in
> arch/*/include/asm/compat.h.
>
> Signed-off-by: Guo Ren <guoren@xxxxxxxxxxxxxxxxx>
> Signed-off-by: Guo Ren <guoren@xxxxxxxxxx>
> Cc: Arnd Bergmann <arnd@xxxxxxxx>
Unfortunately, this one does not look correct to me:
> @@ -116,7 +116,7 @@
> #define F_GETSIG 11 /* for sockets. */
> #endif
>
> -#ifndef CONFIG_64BIT
> +#if !defined(CONFIG_64BIT) || defined(CONFIG_COMPAT)
> #ifndef F_GETLK64
> #define F_GETLK64 12 /* using 'struct flock64' */
> #define F_SETLK64 13
The problem here is that include/uapi/ headers cannot contain checks for
CONFIG_* symbols because those may have different meanings in user space
compared to kernel.
This is a preexisting problem in the header, but I think the change
makes it worse.
With the current behavior, user space will always see the definitions,
unless it happens to have its own definition for CONFIG_64BIT already.
On 64-bit parisc, this has the effect of defining the macros to the
same values as F_SETOWN/F_SETSIG/F_GETSIG, which is potentially
harmful. On MIPS, it uses values that are different from the 32-bit numbers
but are otherwise unused. Everywhere else, we get the definition from
the 32-bit architecture in user space, which will do nothing in the kernel.
The correct check for a uapi header would be to test for
__BITS_PER_LONG==32. We should probably do that here, but
this won't help you move the definitions, and it is a user-visible change
as the incorrect definition will no longer be visible. [Adding Jeff and Bruce
(the flock mainainers) to Cc for additional feedback on this]
For your series, I would suggest just moving the macro definitions to
include/linux/compat.h along with the 'struct compat_flock64'
definition, and leaving the duplicate one in the uapi header unchanged
until we have decided on a solution.
Arnd
