[Devel] Re: [RFC][PATCH 06/16] Define is_global_init()

[sukadev at us.ibm.com]

Dave Hansen [hansendc at us.ibm.com] wrote:
| On Fri, 2007-05-25 at 13:44 -0700, sukadev at us.ibm.com wrote:
| > Dave Hansen [hansendc at us.ibm.com] wrote:
| > | On Thu, 2007-05-24 at 13:24 +0400, Pavel Emelianov wrote:
| > | > > | > +int is_global_init(struct task_struct *tsk)
| > | > > | > +{
| > | > > | > + return (task_active_pid_ns(tsk) == &init_pid_ns && tsk->pid == 1);
| > | > > | 
| > | > > | This can OOPS if you pass arbitrary task to this call...
| > | > > | tsk->nsproxy can already be NULL.
| > | > > 
| > | > > Hmm. You are right. btw, this could be a bisect issue. Patch 9 of uses
| > | > > pid_ns from pid->upid_list and removes nsproxy->pid_ns.
| > | > 
| > | > Yes, but that patch is not good either.
| > | > task_pid(tsk) may become NULL as well and this will oops.
| > | 
| > | Have you reviewed the call paths to make sure this can actually happen
| > | in practice?
| > 
| > task_pid() can be NULL when we are tearing down the task structure in
| > release_task() and in the tiny window between detach_pid() and attach_pid()
| > in de_thread(). 
| > 
| > I think task_pid() is safe as long as it is called for 'current'. (we should
| > probably add some comments)
| 
| If we only call it for "current", then perhaps we should just change the
| function so that it doesn't take any arguments.  That way nobody can
| screw it up.

Well, if the caller can confirm that the tsk passed in to task_pid() is
not exiting, then it is ok to use. We have one such usage in do_fork()
where we use the task struct we just initialized. The task has not yet
been woken up.

| 
| > I will double check my code, but I think all my calls to task_pid() and hence,
| > to task_active_pid_ns() are safe, except for two cases:
| > 
| >         a) is_global_init(). There are a few calls to process other than
| >            current, but not sure if they are a problem.
| > 
| >            For instance in current code, unhandled_signal() checks
| >            tsk->pid == 1 and proceeds to derefernce tsk->sighand.
| > 
| >            If task_pid() is NULL because the task was in release_task(),
| >            then so is tsk->sighand.
| 
| Really?  Are there barriers or locks to make this happen?  Can you be
| sure that compiler or cpu re-ordered code will keep this true?

I don't understand. Here is unhandled_signal() from 2.6.21-mm2.

int unhandled_signal(struct task_struct *tsk, int sig)
{
        if (is_init(tsk))
                return 1;
        if (tsk->ptrace & PT_PTRACED)
                return 0;
        return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
                (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
}

My patch changed the is_init() to is_global_init() but the theory is
that is_global_init() is not safe since it uses task_pid() which
can return NULL if @tsk is exiting.

But if @tsk is exiting, tsk->sighand will also be NULL. So either
the current callers have somehow ensured @tsk is not exiting or
they are risking accessing tsk->sighand.

Not sure how compiler/cpu-reordering changes things. Can you elaborate ?

| 
| >         b) the temporary check I added in check_kill_permissions().
| >            (I need to address Serge's comment here anyway).
| > 
| > To make is_global_init() more efficient and independent of task_pid(),
| > can we steal a bit from task_struct->flags ? Like PF_KSWAPD, and there
| > are unused bits :-)
| 
| It feels to me like we're adding too many hacks on hacks here.  Let's
| define the problem, because it sounds to me like we don't even know what
| it really is.  What is the problem here, again?

We need an interface is_global_init() that tells us if the given process
is /sbin/init (which is the process with pid_t == 1 in init_pid_ns).

Why we need is_global_init(): to allow/deny some facilities (eg: you
cannot send arbitrary signals to the process).

How to implement is_global_init():  One simple/efficient way is to use a
bit in task->flags.

Another way is to:

	- ensure tsk->pid == 1 and
	- active pid ns of @tsk is init_pid_ns.

To check active pid ns of a process we currently need task_pid() which
*may* not be safe for all processes at all times.

A releated interface is is_container_init() which can be defined as
the process with pid_t == 1 in its active pid namespace.
| 
| -- Dave
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers