Re: [RFC PATCH v2 1/4] y2038: add 64bit time_t support in timeval for 32bit architecture

[Bamvor Zhang Jian <bamvor.zhangjian@linaro.org>]

Hi, Arnd

On 07/09/2015 06:26 PM, Arnd Bergmann wrote:
> On Thursday 09 July 2015 17:02:47 Bamvor Zhang Jian wrote:
>> On 07/09/2015 04:09 AM, John Stultz wrote:
>>> On Mon, Jun 29, 2015 at 7:23 AM, Bamvor Zhang Jian
>>> <bamvor.zhangjian@linaro.org> wrote:
>>>> +int get_timeval64(struct timeval64 *tv,
>>>> +                  const struct __kernel_timeval __user *utv)
>>>> +{
>>>> +       struct __kernel_timeval ktv;
>>>> +       int ret;
>>>> +
>>>> +       ret = copy_from_user(&ktv, utv, sizeof(ktv));
>>>> +       if (ret)
>>>> +               return -EFAULT;
>>>> +
>>>> +       tv->tv_sec = ktv.tv_sec;
>>>> +       if (!IS_ENABLED(CONFIG_64BIT)
>>>> +#ifdef CONFIG_COMPAT
>>>> +          || is_compat_task()
>>>> +#endif
>>>
>>> These sorts of ifdefs are to be avoided inside of functions.
>>
>>> Instead, it seems is_compat_task() should be defined to 0 in the
>>> !CONFIG_COMPAT case, so you can avoid the ifdefs and the compiler can
>>> still optimize it out.
>> I add this ifdef because I got compile failure on arm platform. This
>> file do not include the <linux/compat.h> directly. And in arm64,
>> compat.h is included implicitily.
>> So, I am not sure what I should do here. Include <linux/compat.h> in
>> this file directly or add a this check at the beginning of this file?
>>
>> #ifndef is_compat_task
>> #define is_compat_task() (0)
>> #endif
>>
>
> Actually I think we can completely skip this test here: Unlike
> timespec, timeval is defined in a way that always lets user space
> use a 64-bit type for the microsecond portion (suseconds_t tv_usec).
I do not familar with this type. I grep the suseconds_t in glibc, it
seems that suseconds_t(__SUSECONDS_T_TYPE) is defined as
__SYSCALL_SLONG_TYPE which is __SLONGWORD_TYPE(32bit on 32bit
architecture).
> I think we should simplify this case and just assume that user space
> does exactly that, and treat a tv_usec value with a nonzero upper
> half as an error.
>
> I would also keep this function local to the ppdev driver, in order
> to not proliferate this to generic kernel code, but that is something
> we can debate, based on what other drivers need. For core kernel
> code, we should not need a get_timeval64 function because all system
> calls that pass a timeval structure are obsolete and we don't need
> to provide 64-bit time_t variants of them.
Got it.

regards

bamvor
>
> 	Arnd
>