On Mon 2017-02-06 13:51:48, Josh Poimboeuf wrote:

Great catch! Looks good to me.

This assumes that the 'if (klp_target_state == KLP_UNPATCHED)' clause in
klp_try_complete_transition() gets moved to klp_complete_transition() as
you suggested.

quoted

quoted

quoted

quoted

diff --git a/kernel/livepatch/patch.c b/kernel/livepatch/patch.c
index 5efa262..1a77f05 100644
--- a/kernel/livepatch/patch.c
+++ b/kernel/livepatch/patch.c

@@ -29,6 +29,7 @@
 #include <linux/bug.h>
 #include <linux/printk.h>
 #include "patch.h"
+#include "transition.h"
 
 static LIST_HEAD(klp_ops);
 

@@ -54,15 +55,58 @@ static void notrace klp_ftrace_handler(unsigned long ip,
 {
 	struct klp_ops *ops;
 	struct klp_func *func;
+	int patch_state;
 
 	ops = container_of(fops, struct klp_ops, fops);
 
 	rcu_read_lock();
+
 	func = list_first_or_null_rcu(&ops->func_stack, struct klp_func,
 				      stack_node);
+
+	/*
+	 * func should never be NULL because preemption should be disabled here
+	 * and unregister_ftrace_function() does the equivalent of a
+	 * synchronize_sched() before the func_stack removal.
+	 */
+	if (WARN_ON_ONCE(!func))
+		goto unlock;
+
+	/*
+	 * Enforce the order of the ops->func_stack and func->transition reads.
+	 * The corresponding write barrier is in __klp_enable_patch().
+	 */
+	smp_rmb();

I was curious why the comment did not mention __klp_disable_patch().
It was related to the hours of thinking. I would like to avoid this
in the future and add a comment like.

	 * This barrier probably is not needed when the patch is being
	 * disabled. The patch is removed from the stack in
	 * klp_try_complete_transition() and there we need to call
	 * rcu_synchronize() to prevent seeing the patch on the stack
	 * at all.
	 *
	 * Well, it still might be needed to see func->transition
	 * when the patch is removed and the task is migrated. See
	 * the write barrier in __klp_disable_patch().

Agreed, though as you mentioned earlier, there's also the implicit
barrier in klp_update_patch_state(), which would execute first in such a
scenario.  So I think I'll update the barrier comments in
klp_update_patch_state().

You inspired me to a scenario with 3 CPUs:

CPU0			CPU1			CPU2

__klp_disable_patch()

  klp_init_transition()

    func->transition = true

  smp_wmb()

  klp_start_transition()

    set TIF_PATCH_PATCHPENDING

			klp_update_patch_state()

			  task->patch_state
			     = KLP_UNPATCHED

			  smp_mb()

						klp_ftrace_handler()
						  func = list_...

						  smp_rmb() /*needed?*/

						  if (func->transition)

I think this isn't possible.  Remember the comment I added to
klp_update_patch_state():

 * NOTE: If task is not 'current', the caller must ensure the task is inactive.
 * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.

Right now klp_update_patch_state() is only called for current.
klp_ftrace_handler() on CPU2 would be running in the context of a
different task.

I agree that it is impossible with the current code. In fact, I cannot
imagine a way to migrate the task where the barrier would be needed.
The question if we could/should somehow document it. Something like

	* The barrier is not really needed when the patch is being
	* disabled. The value of func->transition would change
	* the result of this handled only after the task is migrated.
	* But the conditions for the migration are very limited
	* and practically include a full barrier, see
	* klp_update_patch_state().

quoted

We need to make sure the CPU3 sees func->transition set. Otherwise,
it would wrongly use the function from the patch.

So, the description might be:

	 * Enforce the order of the ops->func_stack and
	 * func->transition reads when the patch is enabled.
	 * The corresponding write barrier is in __klp_enable_patch().
	 *
	 * Also make sure that func->transition is visible before
	 * TIF_PATCH_PENDING_FLAG is set and the task might get
	 * migrated to KLP_UNPATCHED state. The corresponding
	 * write barrier is in __klp_disable_patch().


By other words, the read barrier here is needed from the same
reason as the write barrier in __klp_disable_patch().

quoted

quoted

quoted

+void klp_reverse_transition(void)
+{
+	unsigned int cpu;
+	struct task_struct *g, *task;
+
+	klp_transition_patch->enabled = !klp_transition_patch->enabled;
+
+	klp_target_state = !klp_target_state;
+
+	/*
+	 * Clear all TIF_PATCH_PENDING flags to prevent races caused by
+	 * klp_update_patch_state() running in parallel with
+	 * klp_start_transition().
+	 */
+	read_lock(&tasklist_lock);
+	for_each_process_thread(g, task)
+		clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
+	read_unlock(&tasklist_lock);
+
+	for_each_possible_cpu(cpu)
+		clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
+
+	/* Let any remaining calls to klp_update_patch_state() complete */
+	synchronize_rcu();
+
+	klp_start_transition();

Hmm, we should not call klp_try_complete_transition() when
klp_start_transition() is called from here. I can't find a safe
way to cancel klp_transition_work() when we own klp_mutex.
It smells with a possible deadlock.

I suggest to move move klp_try_complete_transition() outside
klp_start_transition() and explicitely call it from
 __klp_disable_patch() and __klp_enabled_patch().
This would fix also the problem with immediate patches, see
klp_start_transition().

Agreed.  I'll fix it as you suggest and I'll put the mod_delayed_work()
call in klp_reverse_transition() again.

There is one small catch. The mod_delayed_work() might cause that two
works might be scheduled:

  + one already running that is waiting for the klp_mutex
  + another one scheduled by that mod_delayed_work()

A solution would be to cancel the work from klp_transition_work_fn()
if the transition succeeds.

Another possibility would be to do nothing here. The work is
scheduled very often anyway.

Yes, I think I'll do this, for the sake of simplicity.

Sounds good to me.

I am sorry for the late reply. I am ill and work only limited
time.

Best Regards,
Petr