[PATCH 08/11] nptl: Move cancel state out of cancelhandling

[Florian Weimer]

* Adhemerval Zanella:

> On 26/05/2021 15:20, Florian Weimer wrote:
>> * Adhemerval Zanella via Libc-alpha:
>> 
>>> diff --git a/nptl/descr.h b/nptl/descr.h
>>> index a120365f88..a3084fdf60 100644
>>> --- a/nptl/descr.h
>>> +++ b/nptl/descr.h
>> 
>>> @@ -407,6 +401,10 @@ struct pthread
>>>    /* Used on strsignal.  */
>>>    struct tls_internal_t tls_state;
>>>  
>>> +  /* Thread cancel state (PTHREAD_CANCEL_ENABLE or
>>> +     PTHREAD_CANCEL_DISABLE).  */
>>> +  unsigned char cancelstate;
>> 
>> You could move this into the padding after the c11 flag, I think.
>
> Right, I moved to below c11.  What kind of constraint we have for the
> 'struct pthread' regarding its internal member layout? We have a couple
> of unused fields and might be good to clean them up.

The sanitizers depend on the size to derive the TLS memory area from the
thread pointer.

>> I think there is an implied dependency on PTHREAD_CANCEL_ENABLE == 0 in
>> __tls_init_tp and somewhere in pthread_create.  Maybe add a static
>> assert for PTHREAD_CANCEL_ENABLE == 0?
>
> I think it would be better to add an initialization on __tls_init_tp,
> similar to get_cached_stack.  It would be better if we consolidate
> the 'struct pthread' initialization in a common place.

We depend a lot on zero initialization, though.  Maybe it would be
better to use memset in get_cached_stack.  Unrelated change, of course.

Thanks,
Florian