Re: [PATCH 13/20] signal: Implement force_fatal_sig

From: ebiederm@xmission.com (Eric W. Biederman)
To: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linux-arch <linux-arch@vger.kernel.org>,
	Oleg Nesterov <oleg@redhat.com>,
	Al Viro <viro@zeniv.linux.org.uk>,
	Kees Cook <keescook@chromium.org>
Subject: Re: [PATCH 13/20] signal: Implement force_fatal_sig
Date: Wed, 20 Oct 2021 16:25:46 -0500	[thread overview]
Message-ID: <87ee8fjsmd.fsf@disp2133> (raw)
In-Reply-To: <CAHk-=whe-ixeDp_OgSOsC4H+dWTLDSuNDU2a0sE3p8DapNeCuQ@mail.gmail.com> (Linus Torvalds's message of "Wed, 20 Oct 2021 10:05:21 -1000")

Linus Torvalds <torvalds@linux-foundation.org> writes:

> On Wed, Oct 20, 2021 at 7:45 AM Eric W. Biederman <ebiederm@xmission.com> wrote:
>>
>> Add a simple helper force_fatal_sig that causes a signal to be
>> delivered to a process as if the signal handler was set to SIG_DFL.
>>
>> Reimplement force_sigsegv based upon this new helper.
>
> Can you just make the old force_sigsegv() go away? The odd special
> casing of SIGSEGV was odd to begin with, I think everybody really just
> wanted this new "force_fatal_sig()" and allow any signal - not making
> SIGSEGV special.

There remains the original case that is signal_set up_done
deals with generically.  When sending a signal fails the code
attempts send SIGSEGV and if sending SIGSEGV fails the signal
delivery code terminates the process with SIGSEGV.

To keep dependencies to a minimum and to allow for the possibility of
backports I used "force_sigsegv(SIGSEGV)" instead of
"force_fatal_sig(SIGSEGV)".  I will be happy to add an additional
patch that converts all of those case to force_fatal_sig.

> Also, I think it should set SIGKILL in p->pending.signal or something
> like that - because we want this to trigger fatal_signal_pending(),
> don't we?
>
> Right now fatal_signal_pending() is only true for SIGKILL, I think.

In general when a fatal signal is delivered the function complete_signal
individually delivers SIGKILL to the threads, making
fatal_signal_pending true.

For signals like SIGSYS that generate a coredump that is not currently
true, but in the cases I looked at signal_pending() was enough to
get the code to get_signal(), which dequeues the signals and starts
processing them.

I have a branch queued up for the next merge window that implements per
signal_struct coredumps.  Assuming that does not trigger any user space
regressions I can remove the coredump special case in complete_signal.
That will in turn mean that force_siginfo_to_task does not need to
change sa_handler, blocked or clear SIGNAL_UNKILLABLE, as all of the
cases where that matters today will just wind up with complete_signal
setting a per_thread SIGKILL.

I keep playing with the idea of having fatal_signal_pending depend on a
different flag than the per thread bit for SIGKILL in the per thread
signal set.  That might make it clearer that complete_signal has started
killing the process and it is a start of the killing the process that
triggers fatal_signal_pending.

So far the way fatal_signal_pending works hasn't really been a problem
so I keep putting away ideas of cleaner implementations.

Eric