Re: [RFC] SIGKILL vs. SIGSEGV on late execve() failures

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, Feb 16, 2013 at 01:50:24AM +0000, Al Viro wrote:
> On Fri, Feb 15, 2013 at 04:46:43PM -0800, Shentino wrote:
> > On Fri, Feb 15, 2013 at 4:38 PM, Shentino <shentino@gmail.com> wrote:
> > > On Fri, Feb 15, 2013 at 4:04 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > >> How would you manage to have it masked at that point?  setup_new_exec()
> > >> is inevitable after success of flush_old_exec() and it will do
> > >> flush_signal_handlers() for us.
> > >
> > > I wouldn't know for sure but I read somewhere that even if execve
> > > resets handled signals, it didn't also say that ignored signals were
> > > also reset.  (Source: execve man page.)
> > 
> > Also, apologies for the terminology mix-up.  By masked, I mean that
> > the signal was ignored as directed by userspace a-la signal(SIGSEGV,
> > SIG_IGN).
> > 
> > Plus I *think* that signal ignore masks are preserved across an exec.
> 
> You are correct.  OK, what it means is that we do need that force_sigsegv() -
> either there or in all places in ->load_binary() where we currently have
> send_sig_info(SIGSEGV).  I don't think that it's an urgent hole, but yes,
> it is a bug.  Nice catch.

Arrgh...  OK, I'm a blind idiot.  These places in binfmt_elf.c currently use
force_sig(), not send_sig_info().  Currently == since 2006 when somebody
noticed the problem.  Their counterparts in binfmt_elf_fdpic.c were *not*
noticed.  Anyway, that definitely means we want to do it in a single commit;
the only remaining question is whether we have any problems with somebody
ptracing such execve() and then poking the sucker with ptrace(); that _can_
happen with the current mainline for ELF binaries, so this is not something
new.  I'm low on coffee and about to crash, so I might be missing some
horrible problem with it, but in this case I'm fairly sure that such a problem
would be present in current mainline.