Re: [PATCH, resend] Add x86 32 bit vDSO time function support

[Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>]

On 12-01-2015 13:59, H.J. Lu wrote:
> On Mon, Jan 12, 2015 at 6:49 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Mon, Jan 12, 2015 at 6:28 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>> On Mon, Jan 12, 2015 at 5:51 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>>> On Tue, Jan 6, 2015 at 6:35 AM, Adhemerval Zanella
>>>> <azanella@linux.vnet.ibm.com> wrote:
>>>>> Ping.
>>>>>
>>>>> On 09-12-2014 13:32, Adhemerval Zanella wrote:
>>>>>> Linux 3.15 adds support for clock_gettime, gettimeofday, and time vDSO
>>>>>> (commit id 37c975545ec63320789962bf307f000f08fabd48).  This patch adds
>>>>>> GLIBC supports to use such symbol when they are avaiable.
>>>>>>
>>>>>> Along with x86 vDSO support, this patch cleanup x86_64 code by moving
>>>>>> all common code to x86 common folder.  Only init-first.c is different
>>>>>> between implementations.
>>>>>>
>>>>>> Tested on x86 and x86_64 with a 3.5.0 and a 3.16 kernel.
>>>>>>
>>>>>> --
>>>>>>
>>>>>> 2014-12-09  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
>>>>>>           Stefani Seibold  <stefani@seibold.net>
>>>>>>
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/Makefile [$(subdir) = elf]
>>>>>>       (sysdep_routines): Add dl-vdso here, ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/Makefile [$(subdir) = elf]
>>>>>>       (sysdep_routines): ... not here.
>>>>>>       * sysdeps/unix/sysv/linux/i386/gettimeofday.c: New file: set syscall
>>>>>>       fallback when vDSO is not presented.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/init-first.c (__vdso_clock_gettime):
>>>>>>       Define with libc_hidden_proto/libc_hidden_data_def definitions.
>>>>>>       (_libc_vdso_platform_setup): Rename to __vdso_platform_setup.
>>>>>>       * sysdeps/unix/sysv/linux/i386/init-first.c: New file: likewise.
>>>>>>       * sysdeps/unix/sysv/linux/i386/time.c: New file: likewise.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/bits/libc-vdso.h: Moved to ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/libc-vdso.h: ... here.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/clock_gettime.c: Move to ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/clock_gettime.c: ... here.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/gettimeofday.c: Move to ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/gettimeofday.c: ... here.  Also added
>>>>>>       fallback configurable symbol when vDSO is not available.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/time.c: Move to ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/time.c: ... here.  Also refactored to
>>>>>>       be able to redefine fallback symbol when vDSO is not available.
>>>>>>       * sysdeps/unix/sysv/linux/x86_64/timespec_get.c: Move to ...
>>>>>>       * sysdeps/unix/sysv/linux/x86/timespec_get.c: ... here.
>>>>>>
>>>>>> --
>>>>>>
>>>>>> +# include <dl-vdso.h>
>>>>>> +# include <errno.h>
>>>>>> +
>>>>>> +/* If the vDSO is not available we fall back on the syscall.  */
>>>>>> +static int
>>>>>> +__gettimeofday_syscall (struct timeval *tv, struct timezone *tz)
>>>>>> +{
>>>>>> +  return INLINE_SYSCALL (gettimeofday, 2, tv, tz);
>>>>>> +}
>>>>>> +# define GETTIMEOFAY_FALLBACK (void*) (&__gettimeofday_syscall)
>>>>>> +# undef libc_ifunc_hidden_def
>>>>>> +# define libc_ifunc_hidden_def(name)  \
>>>>>> +  libc_ifunc_hidden_def1 (__GI_##name, __gettimeofday_syscall)
>>>> Why do you define __GI___gettimeofday to __gettimeofday_syscall
>>>> instead of the ifunc version? The x86-64 ifunc  works fine.
>>>>
>>>> --
>>>> H.J.
>>> Why isn't __GI_time IFUNC
>>>
>>> 5822: 000a3360    21 FUNC    LOCAL  DEFAULT   13 __GI_time
>>>
>> The key to call local IFUNC function is that we HAVE to tell GCC
>> it is a global function, not local so that it will go through PLT.
>> Otherwise GCC will perform optimization which will break local
>> IFUNC.  I will see what I can do.
>>
> This requires too many changes with unknown benefits at this stage.  OK with
> copyright year update and sysdeps/unix/sysv/linux/x86_64/x32/init-first.c
> fix.
>
> Thanks.
>
The idea of this patch is to just provide IFUNC calls externally to GLIBC, internal
calls are still by the syscall.  I didn't really dig into the issues, but what I
could tell i386 suffers from the same issue as powerpc32: __gettimeofday is resolved
as a local call and thus no PLT is created, although the IFUNC is still called.

I will change the copyright and the x32 and push, thanks for the review.