KASAN: use-after-free Read in task_is_descendant

[Tetsuo Handa]

Oleg Nesterov wrote:
> On 10/22, Tetsuo Handa wrote:
> > > And again, I do not know how/if yama ensures that child is rcu-protected, perhaps
> > > task_is_descendant() needs to check pid_alive(child) right after rcu_read_lock() ?
> >
> > Since the caller (ptrace() path) called get_task_struct(child), child itself can't be
> > released. Do we still need pid_alive(child) ?
> 
> get_task_struct(child) can only ensure that this task_struct can't be freed.

The report says that it is a use-after-free read at

  walker = rcu_dereference(walker->real_parent);

which means that walker was already released.

> 
> Suppose that this child exits after get_task_struct(), then its real_parent exits
> too and calls call_rcu(delayed_put_task_struct).
> 
> Now, when task_is_descendant() is called, rcu_read_lock() can happen after rcu gp,
> iow child->parent can be already freed/reused/unmapped.
> 
> We need to ensure that child is still protected by RCU.

I wonder whether pid_alive() test helps.

We can get

[   40.620318] parent or walker is dead.
[   40.624146] tracee is dead.

messages using below patch and reproducer.

----------
diff --git a/kernel/ptrace.c b/kernel/ptrace.c
index 99cfddd..0d9d786 100644
--- a/kernel/ptrace.c
+++ b/kernel/ptrace.c
@@ -385,6 +385,7 @@ static int ptrace_attach(struct task_struct *task, long request,
 	if (mutex_lock_interruptible(&task->signal->cred_guard_mutex))
 		goto out;
 
+	schedule_timeout_killable(HZ);
 	task_lock(task);
 	retval = __ptrace_may_access(task, PTRACE_MODE_ATTACH_REALCREDS);
 	task_unlock(task);
diff --git a/security/yama/yama_lsm.c b/security/yama/yama_lsm.c
index ffda91a..a231ec6 100644
--- a/security/yama/yama_lsm.c
+++ b/security/yama/yama_lsm.c
@@ -283,6 +283,11 @@ static int task_is_descendant(struct task_struct *parent,
 		return 0;
 
 	rcu_read_lock();
+	if (!pid_alive(parent) || !pid_alive(walker)) {
+		rcu_read_unlock();
+		printk("parent or walker is dead.\n");
+		return 0;
+	}
 	if (!thread_group_leader(parent))
 		parent = rcu_dereference(parent->group_leader);
 	while (walker->pid > 0) {
@@ -315,6 +320,10 @@ static int ptracer_exception_found(struct task_struct *tracer,
 	bool found = false;
 
 	rcu_read_lock();
+	if (!pid_alive(tracee)) {
+		printk("tracee is dead.\n");
+		goto unlock;
+	}
 
 	/*
 	 * If there's already an active tracing relationship, then make an
----------

----------
#include <unistd.h>
#include <sys/ptrace.h>
#include <poll.h>

int main(int argc, char *argv[])
{
	if (fork() == 0) {
		ptrace(PTRACE_ATTACH, getppid(), NULL, NULL);
		_exit(0);
	}
	poll(NULL, 0, 100);
	return 0;
}
----------

But since "child" has at least one reference, reading "child->real_parent" should
be safe. Therefore, I think that bailing out due to pid_is_alive(child) == false
(like above patch does) cannot avoid this problem...