[COMMITTED] Make sysdeps/generic/internal-signals.h less stubby.

  Three of the functions defined by internal-signals.h were not actually
fulfilling their contracts when the sysdeps/generic version of that
file was used.  Also, the Linux version included several more headers
than the generic version, which is the root cause of a build failure
on Hurd (already addressed in another way, but I think it is proper to
make the headers match).

In principle, __libc_signal_block_all, __libc_signal_block_app, and
__libc_signal_restore_set could be defined in non-system-dependent
code, leaving only __is_internal_signal and __clear_internal_signal to
be defined by a sysdeps header, but the Linux version of
internal-signals.h wants to inline a raw rt_sigprocmask syscall, and I
don't feel like messing with that today.

Tested with cross build to i686-gnu, committed.

	* sysdeps/generic/internal-signals.h: Include signal.h,
	sigsetops.h, and stdbool.h.
	(__libc_signal_block_all): Actually block all signals.
	(__libc_signal_block_app): Likewise.
	(__libc_signal_restore_set): Actually restore the signal mask.
---
 ChangeLog                          |  8 ++++++++
 sysdeps/generic/internal-signals.h | 15 ++++++++++++---
 2 files changed, 20 insertions(+), 3 deletions(-)
  

On 04/04/2018 13:45, Zack Weinberg wrote:
> Three of the functions defined by internal-signals.h were not actually
> fulfilling their contracts when the sysdeps/generic version of that
> file was used.  Also, the Linux version included several more headers
> than the generic version, which is the root cause of a build failure
> on Hurd (already addressed in another way, but I think it is proper to
> make the headers match).
> 
> In principle, __libc_signal_block_all, __libc_signal_block_app, and
> __libc_signal_restore_set could be defined in non-system-dependent
> code, leaving only __is_internal_signal and __clear_internal_signal to
> be defined by a sysdeps header, but the Linux version of
> internal-signals.h wants to inline a raw rt_sigprocmask syscall, and I
> don't feel like messing with that today.

Thanks for catching it.  If I remember correctly, I used an inlined 
rt_sigprocmask for Linux because both sigprocmask and pthread_sigmask 
filter both SIGCANCEL and SIGSETXID. We could use it for 
__libc_signal_block_app, but __libc_signal_block_all and 
__libc_signal_restore_set would require a new symbol.

> 
> Tested with cross build to i686-gnu, committed.
> 
> 	* sysdeps/generic/internal-signals.h: Include signal.h,
> 	sigsetops.h, and stdbool.h.
> 	(__libc_signal_block_all): Actually block all signals.
> 	(__libc_signal_block_app): Likewise.
> 	(__libc_signal_restore_set): Actually restore the signal mask.
> ---
>  ChangeLog                          |  8 ++++++++
>  sysdeps/generic/internal-signals.h | 15 ++++++++++++---
>  2 files changed, 20 insertions(+), 3 deletions(-)
> 
> diff --git a/ChangeLog b/ChangeLog
> index 663e80c195..43e0d6b95f 100644
> --- a/ChangeLog
> +++ b/ChangeLog
> @@ -1,3 +1,11 @@
> +2018-04-04  Zack Weinberg  <zackw@panix.com>
> +
> +	* sysdeps/generic/internal-signals.h: Include signal.h,
> +	sigsetops.h, and stdbool.h.
> +	(__libc_signal_block_all): Actually block all signals.
> +	(__libc_signal_block_app): Likewise.
> +	(__libc_signal_restore_set): Actually restore the signal mask.
> +
>  2018-04-04  Florian Weimer  <fweimer@redhat.com>
>  
>  	inet: Actually build and run tst-deadline.
> diff --git a/sysdeps/generic/internal-signals.h b/sysdeps/generic/internal-signals.h
> index 17ed48c5bf..5282cffce0 100644
> --- a/sysdeps/generic/internal-signals.h
> +++ b/sysdeps/generic/internal-signals.h
> @@ -19,6 +19,10 @@
>  #ifndef __INTERNAL_SIGNALS_H
>  # define __INTERNAL_SIGNALS_H
>  
> +#include <signal.h>
> +#include <sigsetops.h>
> +#include <stdbool.h>
> +
>  static inline bool
>  __is_internal_signal (int sig)
>  {
> @@ -33,20 +37,25 @@ __clear_internal_signals (sigset_t *set)
>  static inline int
>  __libc_signal_block_all (sigset_t *set)
>  {
> -  return 0;
> +  sigset_t allset;
> +  __sigfillset (&allset);
> +  return __sigprocmask (SIG_BLOCK, &allset, set);
>  }
>  
>  static inline int
>  __libc_signal_block_app (sigset_t *set)
>  {
> -  return 0;
> +  sigset_t allset;
> +  __sigfillset (&allset);
> +  __clear_internal_signals (&allset);
> +  return __sigprocmask (SIG_BLOCK, &allset, set);
>  }
>  
>  /* Restore current process signal mask.  */
>  static inline int
>  __libc_signal_restore_set (const sigset_t *set)
>  {
> -  return 0;
> +  return __sigprocmask (SIG_SETMASK, set, NULL);
>  }
>  
>  
>
  

On Wed, Apr 4, 2018 at 1:42 PM, Adhemerval Zanella
<adhemerval.zanella@linaro.org> wrote:
> On 04/04/2018 13:45, Zack Weinberg wrote:
>> Three of the functions defined by internal-signals.h were not actually
>> fulfilling their contracts when the sysdeps/generic version of that
>> file was used.  Also, the Linux version included several more headers
>> than the generic version, which is the root cause of a build failure
>> on Hurd (already addressed in another way, but I think it is proper to
>> make the headers match).
>>
>> In principle, __libc_signal_block_all, __libc_signal_block_app, and
>> __libc_signal_restore_set could be defined in non-system-dependent
>> code, leaving only __is_internal_signal and __clear_internal_signal to
>> be defined by a sysdeps header, but the Linux version of
>> internal-signals.h wants to inline a raw rt_sigprocmask syscall, and I
>> don't feel like messing with that today.
>
> Thanks for catching it.  If I remember correctly, I used an inlined
> rt_sigprocmask for Linux because both sigprocmask and pthread_sigmask
> filter both SIGCANCEL and SIGSETXID. We could use it for
> __libc_signal_block_app, but __libc_signal_block_all and
> __libc_signal_restore_set would require a new symbol.

Yeah, that's what I thought was going on.  I think it's fine the way
it is at least for now.

zw