On 01/05/2020 08:56, Lukasz Majewski wrote:
> Hi Adhemerval,
> 
>> On 26/03/2020 05:06, Lukasz Majewski wrote:
>>> diff --git a/nis/nis_call.c b/nis/nis_call.c
>>> index 92c70e97aa..9c6f62a753 100644
>>> --- a/nis/nis_call.c
>>> +++ b/nis/nis_call.c
>>> @@ -709,7 +709,7 @@ __nisfind_server (const_nis_name name, int
>>> search_parent, nis_error status;
>>>    directory_obj *obj;
>>>    struct timeval now;
>>> -  struct timespec ts;
>>> +  struct __timespec64 ts;
>>>    unsigned int server_used = ~0;
>>>    unsigned int current_ep = ~0;
>>>  
>>> @@ -719,8 +719,8 @@ __nisfind_server (const_nis_name name, int
>>> search_parent, if (*dir != NULL)
>>>      return NIS_SUCCESS;
>>>  
>>> -  __clock_gettime (CLOCK_REALTIME, &ts);
>>> -  TIMESPEC_TO_TIMEVAL (&now, &ts);
>>> +  __clock_gettime64 (CLOCK_REALTIME, &ts);
>>> +  now = valid_timespec64_to_timeval (ts);
>>>  
>>>    if ((flags & NO_CACHE) == 0)
>>>      *dir = nis_server_cache_search (name, search_parent,
>>> &server_used,  
>>
>> I think it would be simpler to just remove the timeval argument on
>> nis_server_cache_search and move the __clock_gettime64 call on the
>> function start. 
> 
> Have I understood you correctly that you recommend removing the "now"
> struct timeval argument and then call explicitly __clock_gettime64 on
> the beginning of nis_server_cache_search function?

Yes, the nis_server_cache_search is a static function used only once
at __nisfind_server.

> 
>>
>> Also, it would require to change nis_server_cache to use a
>> __time64_t for 'expires', otherwise this change won't help in
>> case of a time_t overflow.
>>
> 
> Ok. I will update this. Thanks for pointing this out.
> 
>>
>>> diff --git a/sysdeps/generic/hp-timing.h
>>> b/sysdeps/generic/hp-timing.h index e2d7447212..af9d92f7f7 100644
>>> --- a/sysdeps/generic/hp-timing.h
>>> +++ b/sysdeps/generic/hp-timing.h
>>> @@ -34,8 +34,8 @@ typedef uint64_t hp_timing_t;
>>>     vDSO symbol.  */
>>>  #define HP_TIMING_NOW(var) \
>>>  ({								\
>>> -  struct timespec tv;
>>> 	\
>>> -  __clock_gettime (CLOCK_MONOTONIC, &tv);			\
>>> +  struct __timespec64 tv;
>>> 	\
>>> +  __clock_gettime64 (CLOCK_MONOTONIC, &tv);
>>> \ (var) = (tv.tv_nsec + UINT64_C(1000000000) * tv.tv_sec);	\
>>>  })
>>>    
>>
>> Ok.
>>
>>> diff --git a/sysdeps/generic/memusage.h b/sysdeps/generic/memusage.h
>>> index a111864b0b..91e56d24de 100644
>>> --- a/sysdeps/generic/memusage.h
>>> +++ b/sysdeps/generic/memusage.h
>>> @@ -28,9 +28,9 @@
>>>  #ifndef GETTIME
>>>  # define GETTIME(low,high)
>>> 	   \ {
>>> 			   \
>>> -    struct timespec now;
>>> 	   \
>>> +    struct __timespec64 now;
>>> 		   \ uint64_t usecs;
>>> 			   \
>>> -    clock_gettime (CLOCK_REALTIME, &now);
>>> 	   \
>>> +    __clock_gettime64 (CLOCK_REALTIME, &now);
>>> 		   \ usecs = (uint64_t)now.tv_nsec / 1000 +
>>> (uint64_t)now.tv_sec * 1000000; \ low = usecs & 0xffffffff;
>>> 					   \ high = usecs >>
>>> 32;							   \  
>>
>> Is is the requirement to export __clock_gettime64 as a GLIBC_PRIVATE
>> symbol?
>>
> 
> The __clock_gettime is already exported as GLIBC_PRIVATE at
> ./time/Versions, so I'm following this pattern.
> 
> Moreover, the glibc will not build when __clock_gettime64 is not
> exported.
> 
>> In any case, I think we should try to avoid use internal symbols even
>> for distributed glibc libraries, so I think this change should go
>> once we start to export the clock_gettime64 as default symbol.
> 
> Am I correct that this is a preprocessor macro, which is in the
> exported header?

In fact __clock_gettime is used on other internal libraries and it
is required as is to avoid linknamespace pollution.  So it seems
that __clock_gettime64 should follow the same logic.

It might be misleading that for some ABI __clock_gettime and
for other __clock_gettime64 will be used internally, but it should
be ok nonetheless.