On Wed 2021-10-06 11:04:26, Peter Zijlstra wrote:

On Wed, Oct 06, 2021 at 10:12:17AM +0200, Petr Mladek wrote:

quoted

IMHO, the original solution from v1 was better. We only needed to

It was also terribly broken in other 'fun' ways. See below.

quoted

be careful when updating task->patch_state and clearing
TIF_PATCH_PENDING to avoid the race.

The following might work:

static int klp_check_and_switch_task(struct task_struct *task, void *arg)
{
	int ret;

	/*
	 * Stack is reliable only when the task is not running on any CPU,
	 * except for the task running this code.
	 */
	if (task_curr(task) && task != current) {
		/*
		 * This only succeeds when the task is in NOHZ_FULL user
		 * mode. Such a task might be migrated immediately. We
		 * only need to be careful to set task->patch_state before
		 * clearing TIF_PATCH_PENDING so that the task migrates
		 * itself when entring kernel in the meatime.
		 */
		if (is_ct_user(task)) {
			klp_update_patch_state(task);
			return 0;
		}

		return -EBUSY;
	}

	ret = klp_check_stack(task, arg);
	if (ret)
		return ret;

	/*
	 * The task neither is running on any CPU and nor it can get
	 * running. As a result, the ordering is not important and
	 * barrier is not needed.
	 */
	task->patch_state = klp_target_state;
	clear_tsk_thread_flag(task, TIF_PATCH_PENDING);

	return 0;
}

, where is_ct_user(task) would return true when task is running in
CONTEXT_USER. If I get the context_tracking API correctly then
it might be implemeted the following way:

That's not sufficient, you need to tag the remote task with a ct_work
item to also runs klp_update_patch_state(), otherwise the remote CPU can
enter kernel space between checking is_ct_user() and doing
klp_update_patch_state():

	CPU0				CPU1

					<user>

	if (is_ct_user()) // true
					<kernel-entry>
					  // run some kernel code
	  klp_update_patch_state()
	  *WHOOPSIE*


So it needs to be something like:


	CPU0				CPU1

					<user>

	if (context_tracking_set_cpu_work(task_cpu(), CT_WORK_KLP))

					<kernel-entry>
	  klp_update_patch_state	  klp_update_patch_state()


So that CPU0 and CPU1 race to complete klp_update_patch_state() *before*
any regular (!noinstr) code gets run.

Grr, you are right. I thought that we migrated the task when entering
kernel even before. But it seems that we do it only when leaving
the kernel in exit_to_user_mode_loop().

Which then means it needs to look something like:

noinstr void klp_update_patch_state(struct task_struct *task)
{
	struct thread_info *ti = task_thread_info(task);

	preempt_disable_notrace();
	if (arch_test_bit(TIF_PATCH_PENDING, (unsigned long *)&ti->flags)) {
		/*
		 * Order loads of TIF_PATCH_PENDING vs klp_target_state.
		 * See klp_init_transition().
		 */
		smp_rmb();
		task->patch_state = __READ_ONCE(klp_target_state);
		/*
		 * Concurrent against self; must observe updated
		 * task->patch_state if !TIF_PATCH_PENDING.
		 */
		smp_mb__before_atomic();

IMHO, smp_wmb() should be enough. We are here only when this
CPU set task->patch_state right above. So that CPU running
this code should see the correct task->patch_state.

The read barrier is needed only when @task is entering kernel and
does not see TIF_PATCH_PENDING. It is handled by smp_rmb() in
the "else" branch below.

It is possible that both CPUs see TIF_PATCH_PENDING and both
set task->patch_state. But it should not cause any harm
because they set the same value. Unless something really
crazy happens with the internal CPU busses and caches.

		arch_clear_bit(TIF_PATCH_PENDING, (unsigned long *)&ti->flags);
	} else {
		/*
		 * Concurrent against self, see smp_mb__before_atomic()
		 * above.
		 */
		smp_rmb();

Yeah, this is the counter part against the above smp_wmb().

	}
	preempt_enable_notrace();
}

Now, I am scared to increase my paranoia level and search for even more
possible races. I feel overwhelmed at the moment ;-)

Best Regards,
Petr