Re: [PATCH v2 1/2] linux: Normalize and return timeout on select (BZ #27651)

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

On 09/04/2021 09:38, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> +/* Return true if select modify the timeout to reflect the amount of time
>> +   no slept.  */
>> +extern bool support_select_modify_timeout (void);
>> +
>> +/* Return true if select normalize the timeout input by taking in account
>> +   tv_usec larger than 1000000.  */
>> +extern bool support_select_normalize_timeout (void);
> 
> Maybe use support_select_modifies_timeout and
> support_select_normalizes_timeout?
> 
> Please commit the support/ bits separately.
> 
>> diff --git a/sysdeps/unix/sysv/linux/select.c b/sysdeps/unix/sysv/linux/select.c
>> index 415aa87d3c..d075270ff4 100644
>> --- a/sysdeps/unix/sysv/linux/select.c
>> +++ b/sysdeps/unix/sysv/linux/select.c
>> @@ -33,12 +33,34 @@ int
>>  __select64 (int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds,
>>  	    struct __timeval64 *timeout)
>>  {
>> +  __time64_t s = timeout != NULL ? timeout->tv_sec : 0;
>> +  int32_t us = timeout != NULL ? timeout->tv_usec : 0;
>> +  int32_t ns;
>> +
>> +  if (s < 0 || us < 0)
>> +    return INLINE_SYSCALL_ERROR_RETURN_VALUE (EINVAL);
>> +
>> +  /* Normalize the timeout, as legacy Linux __NR_select and __NR__newselect.
>> +     Different than syscall, it also handle possible overflow.  */
>> +  if (us / USEC_PER_SEC > INT64_MAX - s)
>>      {
>> +      s = INT64_MAX;
>> +      ns = NSEC_PER_SEC - 1;
>>      }
>> +  else
>> +    {
>> +      s += us / USEC_PER_SEC;
>> +      us = us % USEC_PER_SEC;
>> +      ns = us * NSEC_PER_USEC;
>> +    }
>> +
>> +  struct __timespec64 ts64, *pts64 = NULL;
>> +   if (timeout != NULL)
>> +     {
>> +       ts64.tv_sec = s;
>> +       ts64.tv_nsec = ns;
>> +       pts64 = &ts64;
>> +     }
> 
> The overflow handling is slightly inconsistent with the in_time_t_range
> range check below.  Here we use silently saturating arithmetic, below
> it's an error.

In fact on the in_time_t below it needs to check for the normalized 
tv_sec on ts64.tv_sec instead of the input timeout.  I will fix it
and send a newer version without the libsupport bits.