Re: [PATCH v2] Use LFS and 64 bit time for installed programs (BZ #15333)

[Stefan Liebler <stli@linux.ibm.com>]

On 12/01/2022 15:44, Adhemerval Zanella wrote:
> 
> 
> On 12/01/2022 11:13, Stefan Liebler wrote:
>> On 14/12/2021 20:35, DJ Delorie via Libc-alpha wrote:
>> ...
>>>
>>> LGTM
>>> Reviewed-by: DJ Delorie <dj@redhat.com>
>>>
>>
>> Starting with this commit a6d2f948b71adcb5ea395cb04833bc645eab45e6, I
>> get a test fail on s390:
>> FAIL: resolv/tst-p_secstodate
>>
>> Test 0: 0 -> 19700101000000
>> Test 1: 12345 -> 19700101000000
>> test 1 failedTest 2: 999999999 -> 19700101000000
>> test 2 failedTest 3: 2147483647 -> 19700101000000
>> test 3 failedTest 4: 2147483648 -> <overflow>
>> Test 5: 4294967295 -> <overflow>
>>
>> It turns out that there is a type-mismatch of time_t on caller and
>> callee side.
>>
>> in libresolv.so: resolv/res_debug.c: __p_secstodate(u_long secs)
>> time_t clock = secs;
>> struct tm *time = __gmtime_r(&clock, &timebuf);
>>
>> (gdb) p &clock
>> $1 = (time_t *) 0x7ffff1c8
>> (gdb) ptype time_t
>> type = long long
>> (gdb) p	sizeof(time_t)
>> $2 = 8
>> (gdb) x/2xw 0x7ffff1c8
>> 0x7ffff1c8:     0x00000000	0x7fffffff
>>
>> The secs are stored as 8byte long long on stack and the pointer is
>> passed to __gmtime_r in libc.so:
>> time/gmtime.c:
>> /* Provide a 32-bit variant if needed.  */
>> #if __TIMESIZE != 64
>> struct tm *
>> __gmtime_r (const time_t *t, struct tm *tp)
>> {
>>   __time64_t t64 = *t;
>>   return __gmtime64_r (&t64, tp);
>> }
>> #endif
>>
>> (gdb) ptype time_t
>> type = long
>> (gdb) p	sizeof(time_t)
>> $3 = 4
>> (gdb) p         *t
>> $4 = 0
>> (gdb) ptype __time64_t
>> type = long long
>>
>> On libc.so side, time_t is defined as 4byte long, thus the value of t is
>> read as 0x0 instead of 0x7fffffff.
>>
>> resolv/res_debug.c is built with -D_TIME_BITS=64 -D_FILE_OFFSET_BITS=64
>> and time/gmtime.c is built without those defines.
>>
>>
>> @Adhemerval:
>> Can you please have a look?
> 
> 
> libresolv should not be built with the time64 flags, the above patch should fix
> it.
> 
> diff --git a/Makeconfig b/Makeconfig
> index 06f1cca320..9b6fc6b08f 100644
> --- a/Makeconfig
> +++ b/Makeconfig
> @@ -867,7 +867,7 @@ endif
>  +extra-math-flags = $(if $(filter libm,$(in-module)),-fno-math-errno,-fmath-errno)
>  
>  # Use 64 bit time_t support for installed programs
> -installed-modules = nonlib nscd lddlibc4 libresolv ldconfig locale_programs \
> +installed-modules = nonlib nscd lddlibc4 ldconfig locale_programs \
>                     iconvprogs libnss_files libnss_compat libnss_db libnss_hesiod \
>                     libutil libpcprofile libSegFault
>  +extra-time-flags = $(if $(filter $(installed-modules),\
> 
> I have this issue on my backlog since I saw it on powerpc32.  I will check on
> i686 and powerpc and install it.

Thanks for the commit. This fixes the failing test.