Re: [PATCH] x86: do_signal: simplify the TS_RESTORE_SIGMASK logic

[Oleg Nesterov <oleg@redhat.com>]

On 07/13, Matt Fleming wrote:
>
> On Sun, 10 Jul 2011 20:22:03 +0200
> Oleg Nesterov <oleg@redhat.com> wrote:
>
> > 1. do_signal() looks at TS_RESTORE_SIGMASK and calculates the
> >    mask which should be stored in the signal frame, then it
> >    passes "oldset" to the callees, down to setup_rt_frame().
> >
> >    This is ugly, setup_rt_frame() can do this itself and nobody
> >    else needs this sigset_t. Move this code into setup_rt_frame.
> >
> > 2. do_signal() also clears TS_RESTORE_SIGMASK if handle_signal()
> >    succeeds.
> >
> >    We can move this to setup_rt_frame() as well, this avoids the
> >    unnecessary checks and makes the logic more clear.
> >
> > 3. use set_current_blocked() instead of sigprocmask(SIG_SETMASK),
> >    sigprocmask() should be avoided.
>
> Could you please mention commit e6fa16ab "signal: sigprocmask() should
> do retarget_shared_pending()", since it's not immediately obvious in
> this changelog why sigprocmask() should be avoided.

Well, sigprocmask(SIG_SETMASK) is fine from the correctness pov,
it calls set_current_blocked().

sigprocmask() should be avoided because it is strange interface.
It has numeruos callers, but in fact almost all of them could use
set_current_blocked() (ignoring sys_rt_sigprocmask).

Linus suggested to simply kill sigprocmask(). I am not sure, but
at least it shouldn't be abused and its last argument is confusing.

> Reviewed-by: Matt Fleming <matt.fleming@linux.intel.com>

Thanks for looking!

Oleg.