[PATCH 2/2] manual: Document __libc_single_threaded

[Florian Weimer]

* Adhemerval Zanella:

> On 21/05/2020 10:44, Florian Weimer wrote:
>> * Szabolcs Nagy:
>> 
>>> The 05/21/2020 10:15, Adhemerval Zanella wrote:
>>>> On 21/05/2020 10:09, Szabolcs Nagy wrote:
>>>>> The 05/21/2020 09:50, Adhemerval Zanella via Libc-alpha wrote:
>>>>>> On 20/05/2020 15:12, Florian Weimer via Libc-alpha wrote:
>>>>>>
>>>>>>> +@smallexample
>>>>>>> +if (__libc_single_threaded)
>>>>>>> +  atomic_fetch_add (&reference_count, 1, memory_order_relaxed);
>>>>>>> +else
>>>>>>> +  atomic_fetch_add (&reference_count, 1, memory_order_acq_rel);
>>>>>>> +@end smallexample
>>>>>>
>>>>>> Shouldn't the access to __libc_single_threaded be atomic itself
>>>>>> (at least with relaxed semantic)?
>>>>>
>>>>> not if we guarantee that this object can only be
>>>>> written while the process is single threaded.
>>>>>
>>>>> (e.g. an exiting detached thread cannot update it
>>>>> even if only one thread left.. because that may
>>>>> concurrently read it)
>>>>>
>>>>
>>>> OK, so I think we should outline that atomic operations are not required 
>>>> to acess this object and that once __libc_single_threaded is set 0 it will 
>>>> continue to indicate non-single thread even thread are jointed or detached 
>>>> thread finishes.
>>>
>>> what's wrong with pthread_join updating it?
>> 
>> It's tricky do it correctly if there are two remaining threads, one of
>> them the one being joined, the other one a detached thread.  A
>> straightforward implementation merely looking at __nptl_nthreads before
>> returning from pthread_join would not perform the required
>> synchronization on the detached thread exit.
>
> Couldn't we accomplish by making __libc_single_threaded count the total
> number of threads and making pthread_create/pthread_exit/detach exit
> atomically updating it?

We already have __nptl_nthreads as a global thread count, but it is
currently decremented too early (on the exiting thread).  As I tried to
explain, we cannot decrement it on the exiting thread itself because it
would not give us the desired synchronization, particularly not with any
kernel actions that happen afterwards.

Thanks,
Florian