Re: [RFC][PATCH v2 09/11] context_tracking,livepatch: Dont disturb NOHZ_FULL

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 06, 2021 at 12:29:32PM +0200, Petr Mladek wrote:
> On Wed 2021-10-06 11:04:26, Peter Zijlstra wrote:

> > 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().

Yep... :-)

> > 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.

Yes, I think smp_wmb() and smp_mb__before_atomic() are NOPS for all the
same architectures, so that might indeed be a better choice.

> 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.

Right, not our problem :-) Lots would be broken beyond repair in that
case.

> > 		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 ;-)

:-)

Anyway, I still need to figure out how to extract this context tracking
stuff from RCU and not make a giant mess of things, so until that
time....