Re: [Cppcms-users] Getting ready to 1.0.5

[Artyom B. <art...@ya...>]

> 

>>  BUT... I discovered a huge issue with TLS implementation
>>  in the pthreads library that requires workaround.
>> 
>>  If you run pthread_key_create more than 255 times
>>  (even if you call pthread_key_delete and release the keys)
>>  the key creation.
>>  256  is the total number of TLS keys allowed at NetBSD. 
>> 
>> 
>>  However pthread_key_delete does not "free" the spot - it
>>  caused all ICU tests that use TLS for optimization
>>  fail and it may cause other problems that aren't
>>  visible right now.
> 
> That one has been addressed after NetBSD 6 has been branches. The core
> of the issue is that pthread_key_destroy was added as an afterthought
> with some non-trivial implications for how it behaves on existing
> threads.

Destructors are not called if data still exists, use of the key
afterwards leads to undefined behavior.

It is very well defined practically defining the implementation.

  http://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_key_delete.html

All implementation I'm aware of follow this spec.



> Please avoid the term TLS when talking about pthread_key -- it
> is normally used for __thread and the related C++11 feature.
> 

AFAIK Thread local storage is still called TLS regardless implementaion: 
using API like POSIX  pthread_key_create or WinAPI TlsAlloc or
using language based features.

> Joerg
> 

In any case it implies some limits on use of CppCMS in NetBSD other than that
I don't see any other problems. I commited latest fixes so you are welcome to
take a look on them. 



> ------------------------------------------------------------------------------
> _______________________________________________
> Cppcms-users mailing list
> Cpp...@li...
> https://lists.sourceforge.net/lists/listinfo/cppcms-users
>