Re: [PATCH] Use C11 atomics instead atomic_decrement_and_test

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

On 08/09/22 11:06, Wilco Dijkstra wrote:
> Replace atomic_decrement_and_test with atomic_fetch_add_relaxed.
> 
> Passes regress on AArch64.

Some questions below.

> 
> ---
> 
> diff --git a/htl/pt-dealloc.c b/htl/pt-dealloc.c
> index c776e3471dc376977ca7058f0d143f0a842d5799..86bbb3091fbd3758b06294ca705dc774cf8bf0eb 100644
> --- a/htl/pt-dealloc.c
> +++ b/htl/pt-dealloc.c
> @@ -33,7 +33,7 @@ extern pthread_mutex_t __pthread_free_threads_lock;
>  void
>  __pthread_dealloc (struct __pthread *pthread)
>  {
> -  if (!atomic_decrement_and_test (&pthread->nr_refs))
> +  if (atomic_fetch_add_relaxed (&pthread->nr_refs, -1) != 1)
>      return;
>  
>    /* Withdraw this thread from the thread ID lookup table.  */

Ok (and I am not sure why __pthread_create_internal usage does not use atomic at all).

> diff --git a/htl/pt-exit.c b/htl/pt-exit.c
> index f0759c8738e6da4c8e0ba47f6d15720203f6954e..3c0a8c52f34ed64b94ce71a0764512cb76351a3a 100644
> --- a/htl/pt-exit.c
> +++ b/htl/pt-exit.c
> @@ -50,7 +50,7 @@ __pthread_exit (void *status)
>  
>    /* Decrease the number of threads.  We use an atomic operation to
>       make sure that only the last thread calls `exit'.  */
> -  if (atomic_decrement_and_test (&__pthread_total))
> +  if (atomic_fetch_add_relaxed (&__pthread_total, -1) == 1)
>      /* We are the last thread.  */
>      exit (0);
>  
> diff --git a/manual/llio.texi b/manual/llio.texi

I am not sure if MO is suffice here, shouldn't it synchronize with the update
from __pthread_create_internal?

> index 4e6e3fb672bb8ecbb1bb52faf3d70d9b7b33973f..eba9a0e6593f17116b41ff302cfca05846727155 100644
> --- a/manual/llio.texi
> +++ b/manual/llio.texi
> @@ -2614,7 +2614,7 @@ aiocb64}, since the LFS transparently replaces the old interface.
>  @c      free @ascuheap @acsmem
>  @c     libc_thread_freeres
>  @c      libc_thread_subfreeres ok
> -@c     atomic_decrement_and_test ok
> +@c     atomic_fetch_add_relaxed ok
>  @c     td_eventword ok
>  @c     td_eventmask ok
>  @c     atomic_compare_exchange_bool_acq ok

Ok.

> diff --git a/nptl/cond-perf.c b/nptl/cond-perf.c
> index 9c9488e2743f17eb4beeaadf45ef83a38a9f9bf9..baf69e6d34fdb3ffef97b32fb9db422077a0359b 100644
> --- a/nptl/cond-perf.c
> +++ b/nptl/cond-perf.c
> @@ -24,7 +24,7 @@ cons (void *arg)
>  
>    do
>      {
> -      if (atomic_decrement_and_test (&ntogo))
> +      if (atomic_fetch_add_relaxed (&ntogo, -1) == 1)
>  	{
>  	  pthread_mutex_lock (&mut2);
>  	  alldone = true;

Ok, although this code is not used anywhere (neither for testing). Maybe it would be better
to just remove it.

> diff --git a/nptl/pthread_create.c b/nptl/pthread_create.c
> index 2602dba54e872f36bd54955af7587378e5dc3812..ada3fdd4d440c3fdcf4567734a452c8ba4d0fce1 100644
> --- a/nptl/pthread_create.c
> +++ b/nptl/pthread_create.c
> @@ -489,7 +489,7 @@ start_thread (void *arg)
>       the breakpoint reports TD_THR_RUN state rather than TD_THR_ZOMBIE.  */
>    atomic_fetch_or_relaxed (&pd->cancelhandling, EXITING_BITMASK);
>  
> -  if (__glibc_unlikely (atomic_decrement_and_test (&__nptl_nthreads)))
> +  if (__glibc_unlikely (atomic_fetch_add_relaxed (&__nptl_nthreads, -1) == 1))
>      /* This was the last thread.  */
>      exit (0);
>  

Same question from __pthread_total, is relaxed MO suffice here?

> diff --git a/sysdeps/nptl/libc_start_call_main.h b/sysdeps/nptl/libc_start_call_main.h
> index a9e85f2b098dfc6790d719bba9662e428ab237e6..c10a16b2c4a61e74f6bf18396935b87208df4f7e 100644
> --- a/sysdeps/nptl/libc_start_call_main.h
> +++ b/sysdeps/nptl/libc_start_call_main.h
> @@ -65,7 +65,7 @@ __libc_start_call_main (int (*main) (int, char **, char ** MAIN_AUXVEC_DECL),
>        /* One less thread.  Decrement the counter.  If it is zero we
>           terminate the entire process.  */
>        result = 0;
> -      if (! atomic_decrement_and_test (&__nptl_nthreads))
> +      if (atomic_fetch_add_relaxed (&__nptl_nthreads, -1) != 1)
>          /* Not much left to do but to exit the thread, not the process.  */
>  	while (1)
>  	  INTERNAL_SYSCALL_CALL (exit, 0);
>