Re: [PATCH] time: Skip overflow itimer tests on 32-bit systems

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 12/06/2021 06:19, Stafford Horne wrote:
> On Thu, Jun 10, 2021 at 06:38:04AM +0900, Stafford Horne wrote:
>> On Wed, Jun 09, 2021 at 10:50:23AM -0300, Adhemerval Zanella wrote:
>>>
>>>
>>> On 07/06/2021 10:18, Stafford Horne wrote:
>>>> On the port of OpenRISC I am working on and it appears the rv32 port
>>>> we have sets __TIMESIZE == 64 && __WORDSIZE == 32.  This causes the
>>>> size of time_t to be 8 bytes, but the tv_sec in the kernel is still 32-bit
>>>> causing truncation.
>>>>
>>>> The truncations are unavoidable on these systems so skip the
>>>> testing/failures by guarding with __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64.
>>>
>>> Sigh, I was hoping that we won't need to handle this situation (glibc 
>>> support only 64-bit time_t, but kernel still providing some 32-bit 
>>> syscall).
>>>
>>>> ---
>>>>
>>>> I am open to other suggestions, this seemed the most correct to me.
>>>>
>>>> Cc: Adhemerval Zanella <adhemerval.zanella@linaro.org>
>>>> Cc: Alistair Francis <alistair.francis@wdc.com>
>>>>
>>>>  time/tst-itimer.c | 2 ++
>>>>  1 file changed, 2 insertions(+)
>>>>
>>>> diff --git a/time/tst-itimer.c b/time/tst-itimer.c
>>>> index 929c2b74c7..0c99d46d7e 100644
>>>> --- a/time/tst-itimer.c
>>>> +++ b/time/tst-itimer.c
>>>> @@ -89,6 +89,7 @@ do_test (void)
>>>>        TEST_COMPARE (it.it_interval.tv_sec, it_old.it_interval.tv_sec);
>>>>        TEST_COMPARE (it.it_interval.tv_usec, it_old.it_interval.tv_usec);
>>>>  
>>>> +#if __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64
>>>>        if (sizeof (time_t) == 4)
>>>>  	continue;
>>>>  
>>>> @@ -146,6 +147,7 @@ do_test (void)
>>>>  	  TEST_COMPARE (setitimer (timers[i], &it, NULL), -1);
>>>>  	  TEST_COMPARE (errno, EOVERFLOW);
>>>>  	}
>>>> +#endif
>>>>    }
>>>>  
>>>>    {
>>>>
>>>
>>> Instead of disabling, I think it would be better to use
>>> __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64 instead of __time_t sizeof
>>> (so we can still tests the EOVERFLOW):
>>>
>>> diff --git a/time/tst-itimer.c b/time/tst-itimer.c
>>> index 929c2b74c7..bd7d7afe83 100644
>>> --- a/time/tst-itimer.c
>>> +++ b/time/tst-itimer.c
>>> @@ -100,7 +100,7 @@ do_test (void)
>>>
>>>        /* Linux does not provide 64 bit time_t support for getitimer and
>>>          setitimer on architectures with 32 bit time_t support.  */
>>> -      if (sizeof (__time_t) == 8)
>>> +      if (__KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64)
>>>         { 
>>>           TEST_COMPARE (setitimer (timers[i], &it, NULL), 0);
>>>           TEST_COMPARE (setitimer (timers[i], &(struct itimerval) { 0 },
>>> @@ -131,7 +131,7 @@ do_test (void)
>>>        it.it_interval.tv_usec = 20;
>>>        it.it_value.tv_sec = 30;
>>>        it.it_value.tv_usec = 40;
>>> -      if (sizeof (__time_t) == 8)
>>> +      if (__KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64)
>>>         { 
>>>           TEST_COMPARE (setitimer (timers[i], &it, NULL), 0);
>>
>> This looks good to me, I can update to this, test and resend the patch when I
>> get some time.  Probably later tonight.
> 
> I tested this and it exposes an issue in the linux setitimer wrapper.  On my
> platform I get EINVAL instead of EOVERFLOW.
> 
>     FAIL: time/tst-itimer
>     original exit status 1
>     tst-itimer.c:125: numeric comparison failure
>        left: 22 (0x16); from: errno
>       right: 75 (0x4b); from: EOVERFLOW
>     tst-itimer.c:147: numeric comparison failure
>        left: 22 (0x16); from: errno
>       right: 75 (0x4b); from: EOVERFLOW
> 
> 
> It seems this is because sysdeps/unix/sysv/linux/setitimer.c, checks that
> the incoming value is in the range of time_t.  The problem is that
> that we need to fit the value in __int32_t not time_t.  When testing the time_t
> range check does not detect the overflow and setitimer ends up passing a -1 to
> the kernel causing EINVAL.
> 
> I can fix that, as per the patch below, but It will take me some time to audit
> other places this might be an issue.
> 
> 
> 
>   if (! in_time_t_range (new_value->it_interval.tv_sec)
>       || ! in_time_t_range (new_value->it_value.tv_sec))
>     {
>       __set_errno (EOVERFLOW);
>       return -1;
>     }
>   new_value_32.it_interval
>     = valid_timeval64_to_timeval32 (new_value->it_interval);
>   new_value_32.it_value
>     = valid_timeval64_to_timeval32 (new_value->it_value);
> 
> 
> 

Sigh... it seems that openrisc will the only *one* architecture with
64-bit time_t in userland which uses legacy 32-bit kernel ABI. 

> 
> The below patch works for me, but there is probably a better thing to do then
> create a new functrion.
> 
> 
> 
> diff --git a/include/time.h b/include/time.h
> index 4372bfbd96..377a4a45ea 100644
> --- a/include/time.h
> +++ b/include/time.h
> @@ -342,6 +342,14 @@ in_time_t_range (__time64_t t)
>    return s == t;
>  }
>  
> +/* Check whether T fits in a timeval32 (__int32_t).  */
> +static inline bool
> +in_timeval32_range (__time64_t t)
> +{
> +  __int32_t s = t;
> +  return s == t;
> +}
> +

The name is confusing, it is mixing timeval from 'struct timeval' and
time_t.  And there is no need to use __int32_t, we need to use it only
on installed headers to avoid namespace pollution. 

I fact I think it would be better to just change 'in_time_t_range' to
use int32_t internally instead of time_t; I am pretty sure that all
usages assume that sizeof(time_t) == 32.