Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 11 Jul 2018 16:14:56 +0200
From:   Oleg Nesterov <oleg@redhat.com>
To:     "Eric W. Biederman" <ebiederm@xmission.com>
Cc:     Linus Torvalds <torvalds@linux-foundation.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, Wen Yang <wen.yang99@zte.com.cn>,
        majiang <ma.jiang@zte.com.cn>
Subject: Re: [RFC][PATCH 11/11] signal: Ignore all but multi-process signals
 that come in during fork.
Message-ID: <20180711141456.GA6636@redhat.com>
References: <877em2jxyr.fsf_-_@xmission.com>
 <20180711024459.10654-11-ebiederm@xmission.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20180711024459.10654-11-ebiederm@xmission.com>
User-Agent: Mutt/1.5.24 (2015-08-30)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 07/10, Eric W. Biederman wrote:
>
> @@ -1602,6 +1603,20 @@ static __latent_entropy struct task_struct *copy_process(
>  {
>  	int retval;
>  	struct task_struct *p;
> +	unsigned seq;
> +
> +	/*
> +	 * Signals that are delivered to multiple processes need to be
> +	 * delivered to just the parent before the fork or both the
> +	 * parent and the child after the fork.  Cache the multiple
> +	 * process signal sequence number so we can detect any of
> +	 * these signals that happen during the fork.  In the unlikely
> +	 * event a signal comes in while fork is starting and restart
> +	 * fork to handle the signal.
> +	 */
> +	seq = read_seqcount_begin(&current->signal->multi_process_seq);
> +	if (signal_pending(current))
> +		return ERR_PTR(-ERESTARTNOINTR);
>
>  	/*
>  	 * Don't allow sharing the root directory with processes in a different
> @@ -1930,8 +1945,8 @@ static __latent_entropy struct task_struct *copy_process(
>  	 * A fatal signal pending means that current will exit, so the new
>  	 * thread can't slip out of an OOM kill (or normal SIGKILL).
>  	*/
> -	recalc_sigpending();
> -	if (signal_pending(current)) {
> +	if (read_seqcount_retry(&current->signal->multi_process_seq, seq) ||
> +	    fatal_signal_pending(current)) {
>  		retval = -ERESTARTNOINTR;
>  		goto bad_fork_cancel_cgroup;

So once again, I think this is not right, see the discussion on bugzilla.

If signal_pending() == T we simply can't know if copy_process() can succeed or not.
I have already mentioned the races with stop/freeze, but I think there are more.

And in fact I think that the fact that signal_wake_up() helps to avoid the races
with fork() is useful. Say, we could add signal_wake_up() into syscall_regfunc()
and kill syscall_tracepoint_update(). Not that I think this particular change makes
any sense, but it can work.


That is why I tried to sugest another approach. copy_process() should always fail
if signal_pending() == T, just the "real" signal should not disturb the forking
thread unless the signal is fatal or multi-process.

This also makes another difference in multi-threaded case, a signal with a handler
sent to a forking process will be re-targeted to another thread which can handle it;
with your patch this signal will be "blocked" until fork() finishes or until another
thread gets TIF_SIGPENDING. Not that I think this is that important, but still.

Oleg.