[PATCH] Signal semantics for /sbin/init

sukadev at us.ibm.com sukadev at us.ibm.com
Tue Oct 30 14:37:10 PDT 2007


| sukadev at us.ibm.com writes:
| 
| > (This is Oleg's patch with my tweaks to compile, Oleg pls sign-off).
| > ---
| >
| > From: Sukadev Bhattiprolu <sukadev at us.ibm.com>
| > Subject: [PATCH 1/3] Signal semantics for /sbin/init
| >
| > Currently, /sbin/init is protected from unhandled signals by the
| > "current == child_reaper(current)" check in get_signal_to_deliver().
| > This is not enough, we have multiple problems:
| >
| > 	- this doesn't work for multi-threaded inits, and we can't
| > 	  fix this by simply making this check group-wide.
| >
| > 	- /sbin/init and kernel threads are not protected from
| > 	  handle_stop_signal(). Minor problem, but not good and
| > 	  allows to "steal" SIGCONT or change ->signal->flags.
| >
| > 	- /sbin/init is not protected from __group_complete_signal(),
| > 	  sig_fatal() can set SIGNAL_GROUP_EXIT and block exec(), kill
| > 	  sub-threads, set ->group_stop_count, etc.
| >
| > Also, with support for multiple pid namespaces, we need an ability to
| > actually kill the sub-namespace's init from the parent namespace. In
| > this case it is not possible (without painful and intrusive changes)
| > to make the "should we honor this signal" decision on the receiver's
| > side.
| >
| > Hopefully this patch (adds 43 bytes to kernel/signal.o) can solve
| > these problems.
| >
| > Notes:
| >
| > 	- Blocked signals are never ignored, so init still can receive
| > 	  a pending blocked signal after sigprocmask(SIG_UNBLOCK).
| > 	  Easy to fix, but probably we can ignore this issue.
| >
| > 	- this patch allows us to simplify de_thread() playing games
| > 	  with pid_ns->child_reaper.
| >
| > (Side note: the current behaviour of things like force_sig_info_fault()
| >  is not very good, init should not ignore these signals and go to the
| >  endless loop. Exit + panic is imho better, easy to chamge)
| >
| > Oleg.
| >
| > Signed-off-by: Sukadev Bhattiprolu <sukadev at us.ibm.com>
| 
| I kept thinking about the problem and the ignored signal issue
| really disturbed me, because we did not have a clean definition
| on how things are supposed to work.  I do think Oleg was on the right
| track.
| 
| If we make the rule:
| When sending a signal to init.  The presence of a signal handler
| that is neither SIG_IGN nor SIG_DFL allows the signal to be sent to
| init.  If the signal is not sent it is silently dropped, without
| becoming pending.  Further if init specifies it's signal handler as
| SIG_IGN or SIG_DFL all pending signals will be dropped.
| 
| The only noticeable user space difference from todays init
| is that it no longer needs to worry about signals becoming
| pending when it has them marked as SIG_DFL or SIG_IGN, and
| then it blocks them.

Yes, thats a small change in semantics to blocked signals and
may not matter much in practice. 

| 
| This change by making the presence of a signal handler effectively
| a permission check allows us to do all of the work before
| we enqueue the signal.  Which means that we can now allow
| force_sig_info to send signals to init and that panic the kernel
| instead of going into an infinite busy loop taking an exception
| sending a signal and then retaking the same exception.
| 
| This change also makes it possible to easily implement the 
| the desired semantics of /sbin/init for pid namespaces where
| outer processes can kill init but processes inside the pid
| namespace can not.
| 
| Please take a look and tell me what you think.

Overall, it looks good, couple of questions below. Will port my
patches and test it out.

| 
| diff --git a/kernel/signal.c b/kernel/signal.c
| index 4537bdd..79856eb 100644
| --- a/kernel/signal.c
| +++ b/kernel/signal.c
| @@ -546,6 +546,22 @@ static int check_kill_permission(int sig, struct siginfo *info,
|  	return security_task_kill(t, info, sig, 0);
|  }
| 
| +static int sig_available(struct task_struct *tsk, int sig)

Hmm. IMHO, 'available' is not immediately obvious/clear.

| +{
| +	void __user * handler;
| +	int available = 1;
| +
| +	if (likely(!is_global_init(tsk)))
| +		goto out;

With multiple pid namespaces, I guess, this would become

	if (!is_container_init(tsk))
		goto out;

	/* from parent namespace, don't ignore */
	if (!task_in_descendant_ns(tsk))
		goto out;

If this is correct, I have a question below re: do_sigaction.

| +
| +	handler = tsk->sighand->action[sig-1].sa.sa_handler;
| +	available = (handler != SIG_IGN) &&
| +		    (handler != SIG_DFL);

Can we use sig_user_defined() for the above checks ? sig_available()
looks like an extension of sig_user_defined() for init. How about 
sig_init_user_defined() or reverse the logic and use sig_init_ignore()
like Oleg did ?

| +out:
| +	return available;
| +}
| +
| +
|  /* forward decl */
|  static void do_notify_parent_cldstop(struct task_struct *tsk, int why);
| 
| @@ -948,6 +964,9 @@ __group_send_sig_info(int sig, struct siginfo *info, struct task_struct *p)
|  	int ret = 0;
| 
|  	assert_spin_locked(&p->sighand->siglock);
| +	if (!sig_available(p, sig))
| +		return ret;
| +

|  	handle_stop_signal(sig, p);
| 
|  	/* Short-circuit ignored signals.  */
| @@ -1379,6 +1398,11 @@ send_group_sigqueue(int sig, struct sigqueue *q, struct task_struct *p)
|  	read_lock(&tasklist_lock);
|  	/* Since it_lock is held, p->sighand cannot be NULL. */
|  	spin_lock_irqsave(&p->sighand->siglock, flags);
| +	if (!sig_available(p, sig)) {
| +		ret = 1;
| +		goto out;
| +	}
| +
|  	handle_stop_signal(sig, p);
| 
|  	/* Short-circuit ignored signals.  */
	
Hmm. I see now that Oleg's approach would result in the sig_init_ignore()
check being done twice (once in handle_stop_signal() and once in the following
sig_ignored()).

Is that why you don't fold sig_available() into sig_ignored() ? Or, there
other more important/correctness issues as well ?


| @@ -1860,12 +1884,6 @@ relock:
|  		if (sig_kernel_ignore(signr)) /* Default is nothing. */
|  			continue;
| 
| -		/*
| -		 * Global init gets no signals it doesn't want.
| -		 */
| -		if (is_global_init(current))
| -			continue;
| -
|  		if (sig_kernel_stop(signr)) {
|  			/*
|  			 * The default action is to stop all threads in
| @@ -2246,8 +2264,10 @@ static int do_tkill(int tgid, int pid, int sig)
|  		 */
|  		if (!error && sig && p->sighand) {
|  			spin_lock_irq(&p->sighand->siglock);
| -			handle_stop_signal(sig, p);
| -			error = specific_send_sig_info(sig, &info, p);
| +			if (sig_available(p, sig)) {
| +				handle_stop_signal(sig, p);
| +				error = specific_send_sig_info(sig, &info, p);
| +			}
|  			spin_unlock_irq(&p->sighand->siglock);
|  		}
|  	}
| @@ -2336,7 +2356,8 @@ int do_sigaction(int sig, struct k_sigaction *act, struct k_sigaction *oact)
|  		 *   be discarded, whether or not it is blocked"
|  		 */
|  		if (act->sa.sa_handler == SIG_IGN ||
| -		   (act->sa.sa_handler == SIG_DFL && sig_kernel_ignore(sig))) {
| +		   (act->sa.sa_handler == SIG_DFL && sig_kernel_ignore(sig)) ||
| +		   !sig_available(current, sig)) {
|  			struct task_struct *t = current;
|  			sigemptyset(&mask);
|  			sigaddset(&mask, sig);

Not sure about this. Consider that we extend the sig_available() as
I mentioned above for container_init(). Then suppose following sequence
occurs:

	- container-init receives a fatal signal say SIGUSR1, from parent-ns
	  i.e the signal is pending.

	- before processing the pending signal, the container-init sets
	  the handler to SIG_DFL (which is to terminate). 
	  
	  Will we then discard the pending SIGUSR1 even though it was from
	  from parent ns ?



More information about the Containers mailing list