Re: [PATCH v2 1/4] libio: Assume _IO_MTSAFE_IO

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

On 27/04/2022 09:34, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> It is already set by default on all supported architectures and it is
>> an expectation that stdio works on multi-threaded environments.
> 
> So … this cleanup has got stuck in the past because it's actually more
> than just a cleanup.  We actually build most of glibc without libio
> locking.  Looks like misc/ nss/ posix/ have not been covered before.
> 
> Maybe we should just file a bug for the missing locking and fix this
> with this commit?

It seems a better approach, I will check which files as missing support
and open a bug report.

> 
>> index 5af476c48b..c186375c31 100644
>> --- a/sysdeps/nptl/libc-lock.h
>> +++ b/sysdeps/nptl/libc-lock.h
>> @@ -25,7 +25,7 @@
>>  
>>  
>>  /* Mutex type.  */
>> -#if defined _LIBC || defined _IO_MTSAFE_IO
>> +#if defined _LIBC
>>  # if (!IS_IN (libc) && !IS_IN (libpthread)) || !defined _LIBC
>>  typedef struct { pthread_mutex_t mutex; } __libc_lock_recursive_t;
>>  # else
> 
> This doesn't look quite right.  Would we want to compile this
> unconditionally now?

Afaiu using __libc_lock_recursive without _IO_MTSAFE_IO will resulting
in using __libc_lock_recursive_opaque__, which will result in a undefined
type.

Also, I really think we should not tie stdio code with general lock
primitives (as _IO_MTSAFE_IO is doing here).

> 
> There's also some weirdness I can't explain.  I get strange before/after
> symbol differences:
> 
> DIFF eu-readelf -s after strip: nptl/pthread_rwlock_unlock.os
> --- /tmp/Left-ltjeu50v.o        2022-04-27 14:29:39.070599437 +0200
> +++ /tmp/Right-jee20kfm.o       2022-04-27 14:29:39.074599395 +0200
> @@ -1,5 +1,5 @@
>  
> -Symbol table [11] '.symtab' contains 14 entries:
> +Symbol table [11] '.symtab' contains 13 entries:
>   4 local symbols  String table: [12] '.strtab'
>    Num:            Value   Size Type    Bind   Vis          Ndx Name
>      0: 0000000000000000      0 NOTYPE  LOCAL  DEFAULT    UNDEF 
> @@ -12,7 +12,6 @@
>      7: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 ___pthread_rwlock_unlo>
>      8: 0000000000000000      0 NOTYPE  GLOBAL DEFAULT    UNDEF __GI___libc_fatal
>      9: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 __GI___pthread_rwlock_>
> -   10: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 __pthread_rwlock_unlock
> -   11: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 __pthread_rwlock_unloc>
> -   12: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 pthread_rwlock_unlock@>
> -   13: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 pthread_rwlock_unlock@>
> +   10: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 __pthread_rwlock_unloc>
> +   11: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 pthread_rwlock_unlock@>
> +   12: 0000000000000000    463 FUNC    GLOBAL DEFAULT        1 pthread_rwlock_unlock@>
> 
> Can you reproduce this?

I didn't check the symbol generation, I will take a look.