[PATCH][BZ #14782] Do not enable asynchronous cancelation in system

[Ondřej Bílka]

On Sun, Jan 12, 2014 at 12:37:59PM -0500, Carlos O'Donell wrote:
> On 01/12/2014 10:28 AM, Rich Felker wrote:
> > On Sun, Jan 12, 2014 at 01:13:10PM +0100, Ondřej Bílka wrote:
> >> Hi,
> >>
> >> When looking bugs another relatively easy one is that we do not need to
> >> enable async cancellation in system.
> >>
> >> We use cancellation to kill child process and we do not need enable
> >> cancellation until we install handlers to kill a child process. A
> >> cancellation needs to be only enabled in waidpid which already does
> >> that.
> 
> This is not quite correct.
> 
> The child process won't be killed via cancellation (ignore bug 14744 for
> now) because it's another process not a thread.
> 
> We don't use any cancellation to kill the child. The parent is the one 
> which we want to cancel.
> 
> The child won't install the SIGCANCEL handler until it runs
> __pthread_initialize_minimal_internal which may be never if it's not
> MT.
> 
> You are correct in that waitpid is a cancellation point and will check
> to see if this thread was the target of a cancellation request and
> cancel the thread before the function returns.
> 
> Hopefully that clarifies things for you. If you have any questions about
> cancellation please ask them. It's a big mess right now in glibc, and
> I'm happy to share what I know.
> 
Carlos a cancellation handler in sysdeps/unix/sysv/linux/system.c is

/* The cancellation handler.  */
static void
cancel_handler (void *arg)
{
  pid_t child = *(pid_t *) arg;

  INTERNAL_SYSCALL_DECL (err);
  INTERNAL_SYSCALL (kill, err, 2, child, SIGKILL);

  TEMP_FAILURE_RETRY (__waitpid (child, NULL, 0));

How this code does not kill child?