Re: [PATCH] sched/scs: Reset the shadow stack when idle_task_exit

[Woody Lin]

On Tue, Oct 12, 2021 at 6:57 PM Valentin Schneider
<valentin.schneider@xxxxxxx> wrote:
>
> On 12/10/21 18:35, Woody Lin wrote:
> > On Tue, Oct 12, 2021 at 6:00 PM Valentin Schneider
> > <valentin.schneider@xxxxxxx> wrote:
> >>
> >> So AIUI for SCS that works just fine - one thing I'm unclear on is how the
> >> following pops are going to work given the SP reset happens in the middle
> >> of a call stack, but AFAICT that was already the case before I messed about
> >> with init_idle(), so that must already be handled.
> >
> > Hi Valentin,
> >
> > Thanks for the question. The 'scs_task_reset' here resets only the
> > '.thread_info.scs_sp' of the task, so the register (on arm64 it's x18)
> > is still pointing to the same location for popping and storing call
> > frames. The register will be updated to '.thread_info.scs_sp' in
> > '__secondary_switched', which starts a new core and there is no popping
> > after the updating, so it won't introduce an underflow.
> >
>
> I think I got it; __secondary_switched() -> init_cpu_task() -> scs_load()
>
> Thanks!
>
> >>
> >> I'm not familiar enough with KASAN to say whether that
> >> kasan_unpoison_task_stack() should rather happen upon hotplugging the CPU
> >> back (rather than on hot-unplug). If that is the case, then maybe somewhere
> >> around cpu_startup_entry() might work (and then you could bunch these two
> >> "needs to be re-run at init for the idle task" functions into a common
> >> helper).
> >
> > unpoison looks more like an one-time thing to me; the idle tasks will
> > reuse the same stack pages until system resets, so I think we don't need
> > to re-unpoison that during hotplugging as long as it's unpoisoned in
> > 'init_idle'.
> >
>
> I would tend to agree, but was bitten by s390 freeing some memory on
> hot-unplug and re-allocating it upon hotplug:
>
>   6a942f578054 ("s390: preempt: Fix preempt_count initialization")
>
> This makes me doubt whether we can assert the idle task stack pages are
> perennial vs hotplug on all architectures.

I made a quick study on memory-hotplug and seems that only memory contains
nothing other than migratable pages can be unplugged. So process stack
pages should not be a concern for this, since which is an unmovable
memory.

However I don't have a chance to work on a system that enables
memory-hotplug so far, so couldn't verify this assumption further. Guess
we can create a separate thread to clarify this more.

Regards,
Woody

>
> >>
> >> >  }
> >> >
> >> > --
> >> > 2.33.0.882.g93a45727a2-goog