Re: [PATCH 12/23] nptl: Introduce __nptl_deferred_init

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

On 13/05/2021 17:51, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> On 12/05/2021 13:58, Florian Weimer via Libc-alpha wrote:
>>> This function is called once by pthread_create, before spawning
>>> the first thread, and also within pthread_cancel.  Since it sets up
>>> cancellation handlers, place it along with the pthread_cancel
>>> implementation within libc.
>>
>> Why does it need to be initialized by pthread_create as well?  I think
>> moving the initialization to pthread_cancel and only setting once
>> when pthread_cancel is called (and not tying to __libc_single_threaded)
>> is slight better since it avoids the setup of SIGCANCEL for programs
>> that do not use cancellation.
> 
> We need to unblock SIGCANCEL anyway, so that we can send the signal.
> This part can't be deferred until pthread_cancel is called.

Yes, what it seems to be wrong is unblocking it on both pthread_create
*and* on pthread_cancel.

> 
> We could delay installing the SIGCANCEL handler because it's a global
> resource, assuming that unshare (CLONE_SIGHAND) will fail in the kernel.
> 
>> Also, it seems wrong that pthread_cancel.c it the implementation
>> that setups both SIGCANCEL and SIGSETXID.  It also seems wrong
>> the rt_sigprocmask, SIG_UNBLOCK for pthread_cancel.
> 
> Hmm.  Should we have separate initialization functions then?

I think so, it is more logical (pthread_cancel only handles SIGCANCEL).