TIF_NOHZ can escape nonhz mask? (Was: [PATCH v3 6/8] x86: Split syscall_trace_enter into two phases)

[Frederic Weisbecker]

On Thu, Jul 31, 2014 at 06:03:53PM +0200, Oleg Nesterov wrote:
> On 07/31, Frederic Weisbecker wrote:
> >
> > On Wed, Jul 30, 2014 at 07:46:30PM +0200, Oleg Nesterov wrote:
> > > On 07/30, Frederic Weisbecker wrote:
> > > >
> > > > On Tue, Jul 29, 2014 at 07:54:14PM +0200, Oleg Nesterov wrote:
> > > >
> > > > >
> > > > > Looks like, we can kill context_tracking_task_switch() and simply change the
> > > > > "__init" callers of context_tracking_cpu_set() to do set_thread_flag(TIF_NOHZ) ?
> > > > > Then this flag will be propagated by copy_process().
> > > >
> > > > Right, that would be much better. Good catch! context tracking is enabled from
> > > > tick_nohz_init(). This is the init 0 task so the flag should be propagated from there.
> > >
> > > actually init 1 task, but this doesn't matter.
> >
> > Are you sure? It does matter because that would invalidate everything I understood
> > about init/main.c :)
> 
> Sorry for confusion ;)
> 
> > I was convinced that the very first kernel init task is PID 0 then
> > it forks on rest_init() to launch the userspace init with PID 1. Then init/0 becomes the
> > idle task of the boot CPU.
> 
> Yes sure. But context_tracking_cpu_set() is called by init task with PID 1, not
> by "swapper". 

Are you sure? It's called from start_kernel() which is init/0.

But that's not the point. You're right that kernel threads don't care about it.
In fact we only need init/1 to get the flag and also call_usermodehelper(), good point!

Maybe we should just enable it everywhere. Trying to spare the flag on specific tasks will
bring headaches for no much performance win. Kernel threads don't do syscalls and seldom
do exceptions, so they shouldn't hit context tracking that much.

> 
> > > > I still think we need a for_each_process_thread() set as well though because some
> > > > kernel threads may well have been created at this stage already.
> > >
> > > Yes... Or we can add set_thread_flag(TIF_NOHZ) into ____call_usermodehelper().
> >
> > Couldn't there be some other tasks than usermodehelper stuffs at this stage? Like workqueues
> > or random kernel threads?
> 
> Sure, but we do not care. A kernel thread can never return to user space, it
> must never call user_enter/exit().

Yep, thanks for noticing that!

> 
> > > I meant that in the scenario you described above the "global" TIF_NOHZ doesn't
> > > really make a difference, afaics.
> > >
> > > Lets assume that context tracking is only enabled on CPU 1. To simplify,
> > > assume that we have a single usermode task T which sleeps in kernel mode.
> > >
> > > So context_tracking[0].state == context_tracking[1].state == IN_KERNEL.
> > >
> > > T wakes up on CPU_0, returns to user space, calls user_enter(). This sets
> > > context_tracking[0].state = IN_USER but otherwise does nothing else, this
> > > CPU is not tracked and .active is false.
> > >
> > > Right after local_irq_restore() this task can migrate to CPU_1 and finish
> > > its ret-to-usermode path. But since it had already passed user_enter() we
> > > do not change context_tracking[1].state and do not play with rcu/vtime.
> > > (unless this task hits SCHEDULE_USER in asm).
> > >
> > > The same for user_exit() of course.
> >
> > So indeed if context tracking is enabled on CPU 1 and not in CPU 0, we risk
> > such situation where CPU 1 has wrong context tracking.
> 
> OK. To simplify, lets discuss user_enter() only. So, it is actually a nop on
> CPU_0, and CPU_1 can miss it anyway.
> 
> > But global TIF_NOHZ should enforce context tracking everywhere.
> 
> And this is what I can't understand. Lets return to my initial question, why
> we can't change __context_tracking_task_switch()
> 
> 	void __context_tracking_task_switch(struct task_struct *prev,
> 					    struct task_struct *next)
> 	{
> 		if (context_tracking_cpu_is_enabled())
> 			set_tsk_thread_flag(next, TIF_NOHZ);
> 		else
> 			clear_tsk_thread_flag(next, TIF_NOHZ);
> 	}
> 
> ?

Well we can change it to global TIF_NOHZ

> How can the global TIF_NOHZ help?

It avoids that flag swap on task_switch.

> 
> OK, OK, a task can return to usermode on CPU_0, notice TIF_NOHZ, take the
> slow path, and do the "right" thing if it migrates to CPU_1 _before_ it
> comes to user_enter(). But this case is very unlikely, certainly this can't
> explain why do we penalize the untracked CPU's ?

It's rather that CPU 0 calls user_enter() and then migrate to CPU 1 and resume
to userspace.

It's unlikely but possible. I actually observed that very easily on early testing.

And it's a big problem because then the CPU runs in userspace, possibly for a long while
in HPC case, and context tracking thinks it is in kernelspace. As a result, RCU waits
for that CPU to complete grace periods and cputime is accounted to kernelspace instead of
userspace.

It looks like a harmless race but it can have big consequences.

> 
> > And also it's
> > less context switch overhead.
> 
> Why???

Because calling context_switch_task_switch() on every context switch is costly.

> 
> I think I have a blind spot here. Help!
> 
> 
> 
> And of course I can't understand exception_enter/exit(). Not to mention that
> (afaics) "prev_ctx == IN_USER" in exception_exit() can be false positive even
> if we forget that the caller can migrate in between. Just because, once again,
> a tracked CPU can miss user_exit().

You lost me on this. How can a tracked CPU miss user_exit()?

> 
> So, why not
> 
> 	static inline void exception_enter(void)
> 	{
> 		user_exit();
> 	}
> 
> 	static inline void exception_exit(struct pt_regs *regs)
> 	{
> 		if (user_mode(regs))
> 			user_enter();
> 	}

That's how I implemented it first. But then I changed it the way it is now:
6c1e0256fad84a843d915414e4b5973b7443d48d
("context_tracking: Restore correct previous context state on exception exit")

This is again due to the shift between hard and soft userspace boundaries.
user_mode(regs) checks hard boundaries only.

Lets get back to our beloved example:

          CPU 0                                  CPU 1
          ---------------------------------------------

          returning from syscall {
               user_enter();
               exception {
                    exception_enter()
                    PREEMPT!
                    ----------------------->
                                                 //resume exception
                                                   exception_exit();
                                                   return to userspace

Here if we use user_mode(regs) from exception_exit(), we are screwed because
the task is in the dead zone between user_enter() and the actual hard return to
userspace.

user_mode() thinks we are in the kernel, but from the context tracking POV we
are in userspace. So we again risk to run in userspace for an undetermined time
and RCU will think we are in the kernel and disturb CPU 1 with IPIs to report
quiescent states. Also all the time spent in userspace will be accounted as kernelspace.

Yeah the context tracking code gave me a lot of headaches :)