Re: [RFC PATCH v2 4/4] y2038: convert ppdev to 2038 safe

[Arnd Bergmann <arnd@arndb.de>]

On Monday 29 June 2015 22:23:27 Bamvor Zhang Jian wrote:
> diff --git a/include/uapi/linux/ppdev.h b/include/uapi/linux/ppdev.h
> index dc18c5d..d62a47d 100644
> --- a/include/uapi/linux/ppdev.h
> +++ b/include/uapi/linux/ppdev.h
> @@ -74,8 +74,18 @@ struct ppdev_frob_struct {
>  #define PPSETPHASE     _IOW(PP_IOCTL, 0x94, int)
>  
>  /* Set and get port timeout (struct timeval's) */
> -#define PPGETTIME      _IOR(PP_IOCTL, 0x95, struct timeval)
> -#define PPSETTIME      _IOW(PP_IOCTL, 0x96, struct timeval)
> +/* Force application use 64 time_t ioctl */
> +/* TODO: It is an open question about we should use a __xxx_timeval or an
> + * implicit array.
> + * replace struct __kernel_timeval with __s32[4]
> + * replace struct compat_timeval with __s32[2]
> + */
> +#define PPGETTIME      PPGETTIME64
> +#define PPSETTIME      PPSETTIME64
> +#define PPGETTIME64    _IOR(PP_IOCTL, 0x95, struct __kernel_timeval)
> +#define PPSETTIME64    _IOW(PP_IOCTL, 0x96, struct __kernel_timeval)
> +#define PPGETTIME32    _IOR(PP_IOCTL, 0x9c, struct __kernel_compat_timeval)
> +#define PPSETTIME32    _IOW(PP_IOCTL, 0x9d, struct __kernel_compat_timeval)

As commented before, these definitions should probably not be part of the
user-visible header file.

The main reason for using an __s64[2] array instead of struct __kernel_timeval
is to avoid adding __kernel_timeval: 'timeval' is thoroughly deprecated
and we don't want to establish new interfaces with that.

In case of this driver, nobody would ever want to change their user
space to use a 64-bit __kernel_timeval instead of timeval and explicitly
call PPGETTIME64 instead of PPGETTIME, because we are only dealing with
an interval here, and a 32-bit second value is sufficient to represent
that. Instead, the purpose of your patch is to make the kernel cope with
user space that happens to use a 64-bit time_t based definition of 
'struct timeval' and passes that to the ioctl.

	Arnd