Re: __xstat et al. as compat symbols

[Adhemerval Zanella via Libc-alpha <libc-alpha@sourceware.org>]

On 21/10/2020 09:57, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> On 21/10/2020 02:21, Florian Weimer wrote:
>>> * Adhemerval Zanella via Libc-alpha:
>>>
>>>> diff --git a/sysdeps/unix/sysv/linux/alpha/fxstat64.c b/sysdeps/unix/sysv/linux/alpha/fxstat64.c
>>>> index 9d6b8eca32..bcfb55050c 100644
>>>> --- a/sysdeps/unix/sysv/linux/alpha/fxstat64.c
>>>> +++ b/sysdeps/unix/sysv/linux/alpha/fxstat64.c
>>>> @@ -22,9 +22,11 @@
>>>>  #include <kernel_stat.h>
>>>>  #include <sysdep.h>
>>>>  #include <xstatconv.h>
>>>> +#include <shlib-compat.h>
>>>>  
>>>>  /* Get information about the file NAME in BUF.  */
>>>>  int
>>>> +attribute_compat_text_section
>>>>  __fxstat64 (int vers, int fd, struct stat64 *buf)
>>>>  {
>>>>    switch (vers)
>>>> @@ -42,4 +44,12 @@ __fxstat64 (int vers, int fd, struct stat64 *buf)
>>>>        }
>>>>      }
>>>>  }
>>>> -strong_alias (__fxstat64, __fxstat);
>>>> +
>>>> +#if SHLIB_COMPAT(libc, GLIBC_2_0, GLIBC_2_33)
>>>> +strong_alias (__fxstat64, __fxstat_compat)
>>>> +compat_symbol (libc, __fxstat_compat, __fxstat, GLIBC_2_0);
>>>> +#endif
>>>> +
>>>> +#if SHLIB_COMPAT(libc, GLIBC_2_1, GLIBC_2_33)
>>>> +compat_symbol (libc, __fxstat64, __fxstat64, GLIBC_2_1);
>>>> +#endif
>>>
>>> This change breaks a lot of static libraries out there.  Do we really
>>> want to do this immediately, or should we just stop using these symbols
>>> from libc_nonshared.a, and defer the compat symbol treatment to a future
>>> glibc version?
>>
>> This is the similar the libm finite symbols where my understanding was
>> user was supposed to rebuild the library.
> 
> I think the impact of that is a bit narrower because it requires special
> compiler flags.  But I suspect I will ask for reverting that in a year
> as well. 8-/

It is sad that we can't get rid of clunky and not well thought interface,
which kind of issues are you seeing that is making you thinking about
reverting the libm finite symbols?

> 
>> The issue is only for static linking, by stopping using
>> libc_nonshared.a we need to provide the versioned stat symbols anyway.
>> And we still need to handle new ABIs where adding the old xstat
>> symbols does not make sense (such as riscv32 or any new 32-bit
>> architecture).
> 
> You could use a SHLIB_COMPAT conditional without compat_symbol, the two
> are separate.

The problem is with:

  #if SHLIB_COMPAT(libc, GLIBC_2_4, GLIBC_2_33)
  int
  __fxstatat (int vers, int fd, const char *file, struct stat *st, int flag)
  {
  }
  #endif

It won't be built for static and without using SHLIB_COMPAT in this 
case will build the interfaces for ABIs that doe not require it.

> 
>> It would required a new compat_symbol directive to export the symbols
>> *just* for static linking and some building tinkering to just build
>> the xstat symbol for dynamic object and static linking depending of
>> support ABI version.
> 
> The problem is not relying on the contents of libc_nonshared.a, that is
> working still.  It's code that calls __xstat et al. directly via the old
> inline functions in the headers.

Yes, I am aware. I am outlining is I think we would need some thikering
to build for static case that is not easily support with the current
shlib-compat macros we have.

> 
>> If we just defer the compat move, it would add the needless symbols on
>> newer ABIs that we will need to keep supporting indefinitely (even
>> though they won't be used anywhere unless user explicily links against
>> a protected symbol) and we would eventually need to handle the y2038
>> support on old ABIs (with the _TIME_SIZE selection) by either adding
>> proper 64-bit time support for by removing the xstat for good.
> 
> See above, we can (and should) remove it on newer ABIs only.

So now is we are moving in providing what is supposed to be compat symbols 
in the static library case as well. Do we really want to start supporting 
such scenario?